Gmail AI Inbox can turn action items buried in email threads into a suggested to-do list, and Google is now expanding the feature to more paid Google Workspace customers. Google announced the expansion on its official Workspace Updates blog. The rollout began in October, with availability arriving gradually rather than appearing for every eligible account at once.

The feature is designed to help people find the work that needs attention without manually scanning long conversations. Google says AI Inbox identifies action items, deliverables, and deadlines, then summarizes important updates across related email threads.

What Gmail’s AI Inbox does

AI Inbox has two main sections. Suggested to-dos surfaces high-priority items found in incoming email, including things that need a reply or a deadline to track. Topics to catch up on summarizes updates across conversations, such as project changes, travel details, or purchase confirmations.

Users can open the related email for context, take an available action, mark a suggested task as done, or remove a suggestion that is not useful. Google’s AI Inbox help guide says the feature currently draws from messages in Gmail’s Primary tab, not tabs such as Promotions or Social. It does not support attachments, delegated inboxes, or encrypted emails.

AI Inbox is not an autonomous assistant that completes every task. It organizes and summarizes information so users can decide what to do next.

Which Google Workspace plans get AI Inbox?

Google’s October 15 announcement lists these eligible editions for the current rollout:

• Google Workspace Business Starter, Business Standard, and Business Plus
• Google Workspace Enterprise Starter, Enterprise Standard, and Enterprise Plus
• AI Pro for Education
• Frontline Plus

Google says the feature is currently available only to users in the United States whose account language is set to English (US). Rapid Release domains began receiving it on October 12, with visibility rolling out over as many as 15 days. Scheduled Release domains are due to begin on October 26, also with a gradual rollout of up to 15 days.

How it fits into Google’s wider AI push

The AI Inbox expansion is part of a broader effort to make AI useful inside everyday work tools. Google has also added voice-based features to Gmail, Docs, and Keep, while its Gemini tools aim to help users organize work and act on information across apps. Saganote has covered Google Workspace voice features in Gmail, Docs, and Keep and Google’s Gemini agents for everyday business workflows.

The idea overlaps with a wider category of AI tools that can turn requests and messages into follow-up work. For background, see Saganote’s guide to using AI to automate everyday tasks. Gmail’s AI Inbox is more limited than a general-purpose agent: it identifies and organizes potential tasks, while the user remains responsible for reviewing the context and completing the work.

What users should know before relying on it

AI-generated summaries and task suggestions can save time, but they should not replace checking the original message when a deadline, commitment, or business decision matters. Google provides controls to mark suggestions as done, remove them, and submit feedback. Users can also open linked messages to verify why an item appeared.

For organizations, Google says there is no separate administrator control for this specific rollout. Eligible users can access AI Inbox from Gmail on desktop or through the mobile app menu, subject to account eligibility and feature availability.

For now, the practical change is a more organized view of work already contained in email, not a promise that Gmail will automatically carry out those tasks. The rollout is gradual, and users outside the supported US-English availability may not see the feature yet.

Nemotron 3.5 Super VL combines reasoning, tool use, and text, image, document, and video understanding in one model. NVIDIA says it delivers up to five times the throughput of other open models in its size class, based on the company’s evaluations. Details are available in NVIDIA’s official announcement.

Smaller Nemotron models target local AI agents

Nemotron 3.5 Bolt 74B-A9B is designed to fit within 24GB of memory using 2-bit quantization. NVIDIA says it can run local agent tasks such as web research, writing, and code debugging on supported NVIDIA systems. It supports tools including llama.cpp, Ollama, LM Studio, Unsloth, and vLLM.

The smaller Nemotron 3.5 Bolt E4B targets conversational and embedded agents, including local voice assistants and interactive interfaces. NVIDIA lists October 23 as its planned release date, so it should not be described as generally available before then.

The announcement extends NVIDIA’s open-model strategy beyond large server deployments. It follows the company’s earlier Nemotron 3 Embed release for retrieval-augmented generation and agentic retrieval, while other companies are also developing business-focused agents, including Meta’s Muse AI agent for small businesses.

The practical distinction is hardware and workload: Super VL targets complex, multimodal enterprise tasks, while Bolt models are designed for tighter memory budgets and local use. NVIDIA’s performance figures are company-reported benchmarks, so real-world results will depend on the hardware, software stack, and task.

Related Articles

Nvidia Says Its New AI Safety Tools Could Have Stopped the Hugging Face HackNvidia Agrees to Hugging Face Acquisition at $12.9 Billion

An AI agent harness is the software layer that lets an AI model do more than generate a convincing answer. It manages the cycle around the model: providing relevant context, executing approved tool calls, returning results, and deciding whether the task is finished or needs another step. The model supplies the reasoning; the harness makes that reasoning actionable.

That distinction helps explain why two products using similar models can feel very different. One may only suggest code, while another can inspect a repository, edit files, run tests, interpret errors, and try again. The difference is not necessarily a more capable model. It can be the system around it.

TL;DR: What an AI agent harness does

An AI agent harness is the runtime and control layer around an AI model. At a minimum, it manages context, tool execution, and the agent loop. More complete harnesses can also handle memory, planning, permissions, approvals, retries, logging, and long-running tasks. The harness does not make every model correct, but it gives the model a structured way to act, observe the outcome, and adjust.

What is an AI agent harness?

A language model receives input and generates output. In a plain chat, a person often supplies the missing pieces: copying in a file, running the suggested command, reading the error, and asking the model what to try next. An agent harness automates some or all of that handoff.

Suppose a developer asks an AI coding assistant to fix a failing test. The model might propose changing a function, but proposing a change is not the same as editing the file or verifying the fix. A harness can provide the repository context, let the model request a file edit, apply the edit, run the test suite, and return the output for another model turn. The process continues until a stopping condition is met or a human needs to intervene.

This is one practical way to understand the difference between an AI agent and the infrastructure that supports it. An agent is the system performing a task; the harness is the runtime that coordinates its interaction with tools, state, and the environment. In practice, the boundaries can overlap because products use these terms differently.

How an AI agent harness works

A harness typically runs a repeated cycle rather than making one model call and treating the first answer as final.

• Receive the task: Capture the user’s goal and any constraints, such as which files or services the agent may access.
• Prepare context: Supply relevant instructions, conversation history, files, tool descriptions, and previous results.
• Ask the model for the next step: The model may respond with text or request a tool call.
• Validate and execute the action: The harness checks that the requested tool exists and that the action is allowed, then runs it.
• Return the result: Output from the tool, such as a test failure or search result, is added to the next model input.
• Continue or stop: The harness repeats the cycle, asks for human approval, or finishes when the task is complete, blocked, or over its execution budget.

Microsoft’s Agent Framework harness documentation describes this kind of runtime as a composition of model access, tool invocation, conversation state, and approval policies. The exact implementation varies by product, but the underlying pattern is broadly useful.

The key components of an AI agent harness

1. The execution loop

The loop coordinates each model turn and its consequences. It tracks whether the model is asking to use a tool, whether the tool returned a result, and whether another step is needed. Without a clear loop, an application may stop too early, repeat the same failed action, or continue indefinitely.

2. Context and history management

Models only work with the information included in their current input and what their architecture otherwise makes available. The harness decides what information to include: the original request, relevant instructions, recent actions, tool results, and files or records needed for the next decision. Sending everything every time can be expensive and noisy; sending too little can make the agent forget a key constraint. Context management is the work of finding the useful middle ground.

For long tasks, a harness may summarize older turns, keep a compact task state, or store information outside the immediate conversation. That is related to memory, but the terms are not identical: conversation history is a record of what happened, while memory usually means information retained or retrieved to help future decisions.

3. Tools and execution environments

Tools let an agent interact with systems beyond text generation. They may include search, a database query, a calendar, a browser, a shell, or file operations. The model usually requests a tool action; the harness or connected service performs it and returns the result. The model should not be treated as if it directly executed a command just because it described one.

The environment matters. Running a command in an isolated workspace is different from running it on a developer’s main machine or against a production service. A well-designed harness limits which tools are available, validates arguments, and separates low-risk inspection from actions that can delete data, send messages, spend money, or affect other people.

4. State, planning, and task tracking

Multi-step work benefits from a durable record of the goal, completed steps, unresolved work, and important decisions. Some harnesses maintain a task list or plan; others rely on a simpler state machine. Planning can help with complex work, but it is not magic. Plans need to be updated when a tool fails or new information changes the next step.

5. Permissions and human approval

A useful agent needs enough freedom to act, but not unlimited access. The harness can restrict tools by task, require confirmation before sensitive actions, and record what was executed. For example, an agent might be allowed to draft an email automatically but require approval before sending it. This makes the harness part of the application’s safety boundary, not just a convenience wrapper.

6. Feedback, retries, and observability

The result of an action is often the most valuable information in the next turn. A compiler error, failed test, rejected API request, or incomplete search gives the model evidence to work with. The harness must return that evidence accurately, decide whether a retry is sensible, and avoid endless loops. Logs and traces help developers understand which tool ran, what it returned, and where a task went wrong.

A real example: fixing a failing test

Imagine an agent is asked to fix a test in a small Python project. A plain chatbot can suggest a patch, but a harness-backed coding agent can carry the task through a verification cycle:

• Read the project instructions and identify the failing test.
• Inspect the relevant source files and test output.
• Propose and apply a focused edit in the permitted workspace.
• Run the test command and capture its output.
• Send the result back to the model so it can decide whether the change worked.
• Stop when the test passes, the task is blocked, or an approval or execution limit is reached.

The test runner acts as a feedback source. The harness does not guarantee that the patch is correct, but it can replace guesswork with evidence and repeatable checks. OpenAI describes a related engineering approach in its article on building software with Codex and an agent-first environment, where the surrounding tools, repository conventions, and feedback loops are designed to help agents do reliable work.

For a deeper look at the model and product side of this ecosystem, see Saganote’s guide to OpenAI, its models, APIs, agents, and Codex. The important point here is not that every coding product uses the same architecture. It is that a dependable agent needs a way to inspect its work and respond to the results.

AI agent harness vs. agent, framework, and orchestrator

These terms overlap, and there is no single naming convention shared by every vendor. A practical distinction helps, as long as it is treated as a guide rather than a strict industry standard.

An agent framework can include a harness, and a small agent can have a very simple harness. An orchestrator may call agents that each have their own harness. The useful question is not which label a vendor chooses, but which responsibilities the software actually handles. Saganote’s look at DeepSeek Harness and its plugin-based agent runtime is a useful companion example of how a concrete project organizes capabilities around an agent.

The execution environment is another important design choice. In its coverage of Vercel’s Eve and Sandbox approach, Saganote explores how agent infrastructure and sandboxing can be separated from the model itself. That separation matters when a harness needs to run tools or code without giving an agent unrestricted access to a host system.

This is also why a harness is not the same thing as agentic AI. Agentic AI describes a broader approach in which systems pursue goals through multi-step actions. The harness is one implementation layer that can make those actions possible and controllable.

Do you need to build your own harness?

Not always. If the task is a single model response or a small, predictable tool call, a custom loop may be enough. If the agent needs persistent state, multiple tools, approvals, context compaction, retries, or long-running execution, using a maintained framework or harness can save substantial engineering work.

A lightweight custom harness gives a team control over its behavior and dependencies, but the team owns the edge cases. It must handle malformed tool calls, timeouts, repeated failures, permission checks, logging, state recovery, and safe stopping. A prebuilt solution can provide more of these pieces, though developers still need to review its defaults and adapt it to the task.

For a first project, start with the smallest loop that solves a real problem. Add persistent memory, planning, parallel agents, or complex orchestration only when the use case requires them. Saganote’s beginner’s guide to creating an AI agent provides a useful starting point, while AI agent vs. chatbot explains why a multi-step workflow differs from a conversational interface.

Common harness mistakes

• Giving tools too much access: Expose only the tools and data the task needs, and require approval for consequential actions.
• Losing important context: Keep durable task state and include relevant instructions and recent tool results.
• Retrying without a limit: Use retry budgets, timeouts, and clear stop conditions.
• Trusting a tool call without validation: Check arguments, permissions, and expected output before execution.
• Calling a task complete without verification: Use tests, validators, or explicit success criteria where possible.
• Treating logs as optional: Keep enough information to investigate failures without unnecessarily storing sensitive data.

The same principles apply to business automation. A workflow that uses AI to automate everyday tasks still needs to know what it can access, how to handle unexpected results, and when a person must make the final call. More autonomy is only useful when the boundaries are clear.

The harness pattern is not limited to coding. A voice assistant may need to route a request, call a service, manage turn-taking, and recover when speech recognition is uncertain. Saganote’s explainer on how voice AI models work provides context for the model side of that interaction. Similarly, OpenAI Dots and always-on agents in ChatGPT Work illustrate why persistent tasks require more than a single prompt and response.

The bottom line

An AI agent harness is the machinery between a model’s proposed next step and what actually happens in the world. It manages context, executes tools, returns feedback, enforces boundaries, and decides whether to continue. The more complex the task, the more important those details become.

The best harness is not necessarily the most elaborate one. It is the simplest system that can carry out the intended task, verify its work, recover sensibly from failures, and keep consequential actions under appropriate control.

This practical guide walks through using Harvey with Codex, from cloning the repository and configuring the model provider to training the agent on a product, discovering prospects, reviewing messages, and testing reply handling. Start with a small supervised test rather than giving an agent permission to contact a large list immediately.

Commands and features can change. Check the official Harvey repository before running them. The steps here reflect the repository documentation reviewed for this guide.

How an AI sales agent works with Harvey

Harvey is a locally run sales agent built around a repeating workflow. It discovers businesses, profiles their websites, scores prospects against criteria you confirm, drafts email sequences, stages messages in an outbox, and processes replies. Python handles much of the predictable pipeline work, while the configured coding-agent CLI handles tasks that need language understanding.

• Discovery finds businesses from configured sources.
• Profiling reads public company websites and collects observations.
• Qualification compares observations with your ideal customer profile.
• Writing prepares personalized email drafts.
• The outbox holds outgoing messages for review.
• Reply handling classifies incoming messages and queues suggested responses.
• Usage limits and a pause mechanism help control activity.

Harvey runs a heartbeat loop: it wakes up, checks its constraints, chooses the next task, records the result, and sleeps until the next cycle. The priority order is designed to handle replies and pending campaign work before starting more prospecting. That is useful because a sales agent should not keep collecting leads while an existing conversation needs attention.

What you need before starting

Check the current README for the supported Python version and provider requirements. The repository documents using Codex as an optional model provider, along with a supported email provider for sending messages and reading replies.

• Git and a terminal.
• A supported Python version and a virtual environment.
• Codex installed and authenticated. See the official Codex documentation.
• An email provider supported by your installed Harvey version.
• A clear description of your product, target customers, geography, and exclusions.
• Preferably, a dedicated sending domain or mailbox rather than your everyday personal inbox.

Step 1: Clone the repository and open Codex

Clone the official repository, enter its directory, and start Codex from there so it can inspect the project's files.

git clone https://github.com/ethanplusai/harvey.git
cd harvey
codex

Before making changes, ask Codex to read the README and project instructions. Have it explain the current setup rather than guessing from a generic tutorial.

Read the README and project instructions first. Explain how this version of Harvey supports Codex, list the prerequisites, and identify the configuration files I need to review. Do not send email, access a mailbox, or change settings yet. If the README and implementation disagree, show me the relevant files.

This prompt makes Codex inspect the actual checkout before acting. You can then ask it to check whether Python, dependencies, and the Codex CLI are available. Review proposed shell commands before approving them.

Step 2: Install Harvey in a virtual environment

Use the Python version required by the current README. A virtual environment keeps Harvey's dependencies separate from other projects. The following is a typical setup pattern; if the repository specifies a version such as Python 3.13, use that interpreter instead of the system default.

python3 --version
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e .

If installation fails, give Codex the error and ask for the smallest safe fix. Do not let it upgrade or downgrade dependencies without explaining why.

The install failed with this error:
[paste the complete error message]

Inspect the project configuration and explain the likely cause. Propose the smallest safe fix, but do not delete the virtual environment or change dependency versions without explaining why. After I approve, apply the fix and rerun the relevant check.

If the harvey command is not found, confirm that the virtual environment is active and that the editable install completed successfully.

Step 3: Select Codex as the provider

Launching Codex in the terminal does not necessarily configure Harvey to use it. The project has its own provider selection and runtime checks. Ask Codex to locate the exact supported setting in the checked-out version, then review the change before applying it.

Inspect the current provider configuration and show me how this version of Harvey selects Codex as its model provider. Identify the exact file and setting. Do not edit anything until I approve the proposed change. Preserve the default safety checks that disable model tools during Harvey's prompts.

The repository describes an isolated Codex runtime that disables tool-providing features and verifies its effective settings before sending a prompt. Do not weaken those checks just to make setup easier. The model should provide a text decision or draft; Harvey's application code should control workflow actions.

Step 4: Configure email access safely

Harvey needs a supported email provider to send messages and read replies. The repository documents provider-specific options, including Gmail OAuth and SMTP/IMAP in supported versions. Choose one provider and follow the current instructions.

cp .env.example .env

Edit .env locally and fill only the variables required for your provider. Never paste passwords, OAuth secrets, access tokens, or app passwords into Codex prompts or source files. Keep .env out of Git. For Gmail OAuth, complete the browser authorization flow yourself and then verify the provider setting in Harvey's configuration.

Step 5: Teach Harvey about your product

Harvey needs accurate product context to write useful outreach. Its documented training command reads a company website and generates local product knowledge files. Run it against your own public website, then inspect the generated files before using them.

harvey train https://your-company.example

Replace the example URL with your company's website. Check the generated description, benefits, pricing, use cases, objections, and competitor claims. Remove inaccuracies and confidential details. Do not let an agent invent product capabilities, customer results, or offers.

Next, review harvey.yaml or the local override supported by your version. The README describes settings for the sender persona, product, ideal customer profile, email provider, approval requirement, send limits, usage budget, heartbeat interval, and quiet hours.

# Illustrative configuration shape only.
# Check exact field names and accepted values in your installed version.
persona:
  name: "Sender name"
  company: "Your company"
  role: "Your role"
product:
  name: "Product name"
  description: "What it does and who it helps"
icp:
  industries: ["Target industry"]
  geography: ["Target region"]
channels:
  email:
    require_approval: true
    max_daily_sends: 10

This snippet illustrates the settings to think about; it is not a complete drop-in configuration. Preserve the existing schema and keep approval enabled with a low send ceiling while testing.

Step 6: Define the ideal customer profile

An ideal customer profile (ICP) describes the companies most likely to benefit from your product. “Any small business” is too broad. Define the industry, geography, company size, business model, and observable signals that make a company relevant. Add exclusions too, such as industries you do not serve or companies that are too small to use the product.

Harvey includes a signal catalog and asks users to review which signals it may collect. Start with free signals where possible, and inspect each signal's description, coverage, and cost before enabling paid sources.

harvey signals
harvey signals --confirm free
harvey discover --estimate
harvey discover

Check the installed version's help and README before running these commands. Discovery providers have different coverage and prices. A free map-based source can be useful for testing local businesses but may miss companies without a mapped storefront. Review the estimate and provider before starting any paid discovery run.

After discovery, profile the businesses and inspect the results. Begin with a small cohort. Reject irrelevant companies and refine the ICP rather than trying to fix poor targeting with generic email copy.

Step 7: Generate and review email drafts

Once prospects are qualified, Harvey can prepare a short sequence based on product knowledge and its email-writing skills. Useful personalization points to a specific, verifiable observation about the business. It should not rely on a guessed pain point or a generic compliment.

Review each draft for accuracy, relevance, tone, recipient identity, merge fields, and a clear but modest next step. A first email should open a conversation rather than imply that the recipient has already expressed interest.

harvey outbox

Use the outbox or dashboard to review queued messages. Approve only drafts that pass your checks; reject or edit the rest. Harvey documents a deterministic pre-send gate for problems such as unrendered merge fields, banned phrases, excessive length or links, invalid addresses, and recipient-record mismatches. These checks help, but they do not replace human review.

Step 8: Run the agent and inspect the dashboard

Use the documented dashboard and status commands after configuration. The dashboard is designed to show activity, prospects, campaigns, conversations, outbox decisions, and usage information. Understand how to stop the process before starting a loop.

harvey dashboard
harvey status
harvey usage
harvey run

The repository describes a local dashboard commonly available at http://localhost:5555. Use the address reported by your installed version if it differs. Keep the terminal open while the process runs, and learn the pause command before starting a campaign.

harvey sending pause

Pause sending if the wrong audience is selected, an unexpected message appears, replies are mishandled, or activity exceeds the intended limits. Verify the command syntax in your installed version before relying on it during an incident.

Step 9: Test reply handling

Harvey's documented reply handler checks incoming messages, deduplicates them, classifies intent, advances conversation stages, and queues suggested responses through the approval process. It also describes bounce handling and cancellation of queued follow-ups.

Test the flow with controlled messages before relying on it. Confirm that Harvey distinguishes a positive response, an objection, a stop request, an out-of-office message, and a bounce. An opt-out must stop further sales follow-ups. Check that a bounce invalidates the address and cancels queued messages where the implementation says it should.

For a meeting request, confirm the details yourself. The repository version reviewed lists calendar integration and automatic booking as roadmap work, so do not assume the agent can book meetings automatically without checking the current project status.

Useful Codex prompts for setup and review

Codex is most useful as a repository-aware assistant. Ask it to explain the implementation, diagnose a failure, or review a configuration. Make one change at a time and request the exact checks it ran.

Explain the Harvey setup flow from the current README. Separate required steps from optional integrations, and identify the relevant repository files.

Review my Harvey configuration for an initial test run. Check the ICP, provider selection, usage budget, quiet hours, and email approval setting. Do not print secrets or modify files. List unsafe or incomplete settings.

Review the proposed campaign drafts in the local outbox. Flag unsupported claims, unverifiable personalization, incorrect merge fields, duplicate recipients, and messages that should not be sent. Do not approve or send anything.

Run relevant non-sending tests for the changes I made. Do not access the mailbox, send email, approve the outbox, or disable safety checks. Summarize the commands run and their results.

If Codex suggests weakening runtime isolation, enabling model tools, disabling approval, or increasing the send limit, ask it to explain the risk and show the exact code path before accepting the change. A successful installation is not proof that the agent is safe to run unattended.

Email safety and deliverability

Cold-email rules vary by jurisdiction and recipient type. Check the laws that apply to the sender and recipient, identify the sender honestly, provide a working opt-out method, and honor opt-outs promptly. This guide is not legal advice; seek qualified counsel before sending at scale or contacting regulated industries.

• Use a dedicated sending domain and configure SPF, DKIM, and DMARC.
• Warm up a new mailbox gradually and begin with low volume.
• Verify addresses rather than sending to guesses.
• Monitor bounces, complaints, and replies.
• Maintain a suppression list and stop follow-ups after an opt-out.
• Keep human approval enabled while testing.
• Store credentials in local secret files or supported OAuth flows, never in Git.
• Treat scraped websites and incoming email as untrusted content, not instructions for the model.
• Do not automate LinkedIn browser actions without checking its terms and account risks. Harvey's repository warns that this can violate LinkedIn's terms and lead to restrictions.

Sending limits and a pause mechanism reduce risk, but they cannot guarantee inbox placement or legal compliance. The safest first campaign is small enough for a person to inspect every recipient, draft, and reply.

Troubleshooting common problems

A safer way to start

Make Harvey prove each stage independently: install it, configure Codex, train it on your product, define a narrow ICP, inspect a small discovery run, review drafts, and test reply handling with controlled messages. Expand only after each stage produces results you can verify.

Codex can help interpret the repository and troubleshoot setup, but it should not make business decisions invisible. Keep prospect criteria explicit, personalization traceable to real observations, the outbox reviewable, and the stop mechanism easy to use. Those controls matter more than maximizing the number of emails an agent can send.

Related Articles

Claude Code Mods Explained: How to Customize AI Coding WorkflowsWhat Does OpenAI Do? A Complete Guide to ChatGPT, Models, APIs, Agents, and CodexClaude Code Can Now Build and Test iOS Apps Directly in the iOS Simulator

Google Chrome vs. Chromium can look like a choice between two versions of the same browser because they share much of the same code. The key difference is that Chromium is the open-source project and browser codebase, while Google Chrome is Google's finished browser product built on that foundation. Google adds its own branding, distribution, update mechanisms, and integrations to Chrome.

For most people who want a browser that installs easily and keeps itself updated, Chrome is the straightforward option. Chromium is worth considering when open-source code, browser development, or testing matters more than a turnkey setup. It is not automatically faster, safer, or more private simply because its source code is available.

What Is Google Chrome?

Google Chrome is a web browser developed and distributed by Google. It uses Chromium as its technical foundation and adds product features and services that make it ready for everyday use. Users can install Chrome, sign in with a Google Account if they choose, and sync supported browser data across devices.

Chrome's Google integrations can include syncing bookmarks, passwords, settings, history, and open tabs when the relevant options are enabled. It also uses Google's update system to deliver browser updates and security fixes. These services are convenient, but they are part of the product experience that users should understand when considering privacy and data settings.

Chrome is available for major desktop and mobile platforms, although the way it is installed and updated depends on the operating system. On Windows and macOS, Google's updater helps keep the browser current. Mobile updates are managed through the relevant app store.

Official references: Google Chrome Help: how Chrome updates work and Google Chrome Help: sign-in and sync.

What Is Chromium?

Chromium is an open-source browser project that provides the code behind Google Chrome and other Chromium-based browsers. The project publishes source code, technical documentation, and ways for developers to build, test, and contribute to the browser.

Because the code is available, developers can compile Chromium, inspect its implementation, modify it, or use it as a starting point for another browser. That flexibility is useful for software development and testing. It also means that the experience depends on which Chromium build a person downloads and who maintains it.

There is no single consumer experience that applies to every Chromium download. A distribution may package updates differently, include or omit certain services, or make its own choices about features and supported media. Before installing a build, check its publisher, update process, and documentation.

The official Chromium project site describes Chromium as the open-source project behind Google Chrome.

Google Chrome vs. Chromium: The Main Differences

Updates and Security: Which Is Safer?

A browser's security depends on more than its name or whether its source code is open. It also depends on how quickly security fixes reach the installed version, whether updates are applied reliably, and how the browser is configured.

Chrome uses Google's update system on supported desktop platforms to check for and install updates, including security fixes. Chromium builds may also receive frequent updates, but the process depends on the package or distributor. A build that is not maintained or updated promptly can leave users exposed to bugs that have already been fixed elsewhere.

For everyday use, Chrome is the simpler default if the alternative is an unfamiliar Chromium download with unclear maintenance. A Chromium build from a reputable operating-system repository or a known project can also be a reasonable choice, but users should verify how it receives security updates.

Is Chromium More Private Than Chrome?

Not automatically. Chromium's open-source code can be inspected, but the privacy behavior of an installed browser also depends on the build, its defaults, included components, and the services a user enables. Source availability alone does not prove that a particular build collects less data or has safer settings.

Chrome includes Google services and lets users choose whether to sign in and which supported information to sync. Google's documentation explains that Chrome Sync can store selected browser information in a Google Account so it can be used on other devices. Users can review sync controls and account activity settings to understand what is enabled.

Chromium builds may omit some Google-specific services, but the details vary. A third-party build could add its own services or change settings, so the word “Chromium” is not a complete privacy policy. Read the distributor's documentation and inspect the settings of the exact build you plan to use.

If privacy is the priority, start with practical questions: Who maintains the browser? How quickly are security updates released? Which telemetry or background services are enabled? Can those settings be disabled? What data is synced, and to which account or provider?

Extensions, Websites, and Media Compatibility

Chrome and Chromium share the same broad browser foundation, so many websites and Chromium-compatible extensions behave similarly in both. However, compatibility is not guaranteed to be identical across every build. Extension availability can depend on the browser's support and distribution, while media playback can depend on bundled codecs, digital rights management components, and operating-system support.

If a website or streaming service does not work in a Chromium build, the cause may be a missing component or a difference in packaging rather than the Chromium engine itself. Chrome is generally less work for users who want the standard Google-supported experience. Developers may prefer Chromium when they need to test how a site behaves in a browser build close to the upstream project.

For a broader look at Google's current browser features, see Saganote's guide to Chrome's vertical tabs, PDF annotations, split view, and Google Wallet autofill.

Performance: Is Chromium Faster Than Chrome?

There is no universal rule that Chromium is faster. Both share much of the same browser engine, and real-world speed depends on the build, operating system, hardware, extensions, open tabs, website code, and background activity.

A stripped-down Chromium build may omit some services, but that does not guarantee a noticeable speed improvement. Another build may use different defaults or lag behind current optimizations. A meaningful comparison should use current versions on the same computer, with the same websites, extensions, settings, and workload.

If Chrome already runs well on a device, switching to Chromium solely for a presumed speed boost is unlikely to be a reliable strategy. Disabling unnecessary extensions, keeping the browser updated, and reducing resource-heavy tabs are more practical first steps.

Which Browser Should You Use?

Choose Google Chrome if...

• You want a straightforward installation and an established update process.
• You use Google Account sync for bookmarks, passwords, settings, or tabs across devices.
• You want Google's standard browser experience with minimal setup.
• You rely on websites or media services that expect Chrome's included components.

Consider Chromium if...

• You develop websites or browser software and need access to the upstream project.
• You want to inspect, compile, or modify open-source browser code.
• You use a trusted Linux distribution or project that maintains a Chromium package.
• You are comfortable checking update schedules, included components, and configuration differences.

If your main goal is privacy rather than development, compare the actual privacy settings and policies of the specific browsers available to you. Chromium is a codebase, not a guarantee that every build has the same privacy properties.

Chromium also sits behind a wider group of browsers with their own features and policies. For example, AI features are becoming part of the browser experience: Saganote has covered Google's Gemini Spark integration in Chrome and Claude in Chrome's browser agent. Those products illustrate why choosing a browser can involve more than the underlying rendering engine.

The Bottom Line

Google Chrome is the practical choice for most people who want a ready-to-use browser with Google-managed updates and optional account sync. Chromium is more useful when access to the open-source project, development, or a trusted distribution's build matters. Neither browser is automatically faster or more private in every situation, so the best choice depends on the build, update process, services, and settings that matter to you.

ChatGPT Chrome tabs can now provide context for conversations, while OpenAI's desktop app can use supported browsers to carry out website tasks. Rather than asking everyone to switch to a separate AI browser, OpenAI is making the existing browser part of its assistant workflow. Rather than asking everyone to switch to a separate AI browser, OpenAI is making the existing browser part of its assistant workflow.

The change follows a broader shift in OpenAI's browser strategy. The company had introduced ChatGPT Atlas as a standalone AI browser, but recent reporting says Atlas is being discontinued. The newer approach puts more emphasis on connecting ChatGPT to Chrome and other browsers. OpenAI's official browser extension guide describes support for Chrome, Edge, Brave, Opera, and Vivaldi, although the available features differ by browser and workspace.

What ChatGPT can do with open browser tabs

The extension can bring a page or selected text into a ChatGPT conversation, and users can mention open tabs when asking a question. That gives ChatGPT context from pages already open in the browser instead of requiring users to copy and paste every detail manually.

For example, a user comparing several products could ask ChatGPT to review multiple product pages and summarize differences. Someone researching software plans could ask it to collect limits and features from open tabs. The usefulness depends on the pages being accessible and the task being clear, but the core idea is to turn browser research into a task ChatGPT can help organize.

OpenAI's documentation also describes browser control from the ChatGPT desktop app. Users can choose a supported browser in a Work or Codex chat, then ask ChatGPT to work with a website where they are already signed in. Depending on permissions and availability, that can include navigating pages, interacting with controls, filling forms, and downloading files.

The strategy is to bring AI into existing browsers

A separate AI browser asks people to change where they browse. An extension-based approach meets people inside a familiar browser and can use their current tabs or signed-in sessions when permission is granted. That makes the browser less of a destination and more of a workspace that an AI assistant can help operate.

OpenAI is not alone in pursuing this model. Google is also bringing AI task features into Chrome, including its Gemini browser efforts. Saganote has covered Google's Gemini Spark Chrome integration, which reflects the same broader push to make browser-based research and routine tasks available through an AI assistant.

Anthropic is taking a related route with Claude in Chrome, which brings Claude into browser workflows. Meanwhile, developer-focused tools are also adding browser capabilities, as explored in Saganote's coverage of Claude Code's built-in browser. These products differ in scope, but all point toward assistants that can use the same interfaces people already work in.

Website access still needs boundaries

Browser control is more sensitive than asking a chatbot to summarize text. A browser may contain private messages, work dashboards, customer records, or accounts that can make changes. Giving an assistant access to a signed-in session can expose information or allow actions that affect real accounts.

OpenAI's documentation says users are prompted before ChatGPT interacts with a new website by default, and website permissions can be managed. Workspace administrators may also restrict browser use, file transfers, or other capabilities. Availability depends on the desktop app, browser, rollout, and workspace settings.

These controls matter because an AI agent can misread a page, follow misleading instructions embedded in site content, or take the wrong action. Permission prompts reduce some risks, but they do not replace human review. For purchases, account changes, messages, or submissions, users should confirm the details before allowing an action to complete.

What this means for ChatGPT users

The practical change is that ChatGPT can work with browser context rather than only answering questions about text a user manually provides. That may make multi-page research, comparison tasks, and repetitive website workflows easier to delegate. It also places more responsibility on users to decide which websites the assistant can access and what it should be allowed to do.

For people who want to try it, OpenAI's official browser extension documentation explains setup, supported browsers, tab mentions, side chat, and website permissions. The desktop app and workspace must support the relevant features, so the exact experience may not be identical for every account.

OpenAI's browser direction is therefore less about replacing Chrome outright and more about making ChatGPT a layer that can read and act within it. Whether that becomes a routine way to work will depend on how reliably it handles real websites and how clearly users can control its access.

The update uses build number CP41.260831.019 and is dated October 16, 2026. Google’s system images cover supported devices from the Pixel 6a through the Pixel 11 series, including Pixel Fold models and the Pixel Tablet, as well as the Android Emulator.

There is one caveat for the newest foldable: Google says Pixel 11 Pro Fold owners may need to enroll in Face Unlock again for it to work properly after installing Beta 7.1.

The previous release, Android 17 QPR2 Beta 7, addressed excessive idle battery drain linked to Flip to Shh and failed HTTPS downloads. Earlier testing also brought a range of Pixel fixes in Android 17 QPR2 Beta 4.

Beta 7.1 is a maintenance update rather than a feature-focused release. As with other preview builds, users should expect potential bugs and check Google’s release notes for device-specific details before installing.

The result puts Jev at the top of Vercel’s September model-launch rankings for reach. It does not mean Jev is the most-used AI model across the industry. Vercel’s figures cover anonymized traffic routed through its own gateway, not every provider’s direct API usage.

What Jev does differently

Jev is a decision model from TypeSafe AI. Rather than focusing on open-ended text generation, it evaluates application state against defined questions and returns structured results that software can use. Depending on the task, an application can ask it to choose between options, score a case, or evaluate a yes-or-no question.

That makes Jev a potential fit for bounded decisions inside an application, such as classifying a support request, routing a task, scoring risk, or flagging an action for review. It is not a general replacement for a language model that writes explanations, generates code, or handles broad conversations.

For more background on its design, see Saganote’s guide to what Jev AI is and how TypeSafe’s System One decision model works.

What Vercel’s adoption numbers show

Vercel reports that Jev reached 35.3% of paid teams during its first 14 days on AI Gateway. Claude Sonnet 5.5 ranked second at 18.1%, followed by OpenAI’s GPT-6 Luna at 14.9%. Jev also received the most requests of any model in September, despite launching partway through the month.

Adoption is only one measure of a model’s performance in the market. Vercel says Jev accounted for 1% of September gateway tokens and 0.15% of gateway spending for the month. In its two-week launch window, Jev’s spending share was 0.27%. The small spend share fits the model’s narrower decision-making role and does not, by itself, establish whether it is better or worse than a larger model.

Return usage provides another signal. Of the teams that tried Jev in its first week, 55.6% used it again in the second week, according to Vercel. That was the fifth-highest retention rate among the September launches included in the report.

Where developers could use a decision model

A developer does not need to rebuild an entire AI application around Jev to test it. A narrower approach is to identify one decision currently handled by a general-purpose model, define the allowed outcomes, and compare Jev against the existing implementation.

• Request routing: choose a destination or queue from a fixed set of options.
• Classification: assign a message or record to a known category.
• Risk checks: flag a proposed action for additional review.
• Workflow decisions: select a next step based on application state.
• Human escalation: identify cases that should not proceed automatically.

These are possible use cases for decision models, not a guarantee that Jev will perform well on every task. Teams should test it against representative examples and compare accuracy, latency, operating cost, and the consequences of incorrect decisions.

Saganote’s practical guide to using Jev AI walks through defining a decision, supplying state, checking results, and connecting them to application logic.

Fast adoption does not prove a model is right for every task

The Vercel data shows that many paid gateway teams tried Jev quickly. It does not reveal why every team adopted it, and it cannot prove that Jev will outperform a general-purpose model on a particular application. Teams may be testing it for a narrow workload, comparing alternatives, or using it alongside other models.

For production use, model output should remain one input to the application’s decision process. Code should still enforce permissions, validate arguments, and control tool execution. For agent workflows, see Saganote’s guide to building AI agents with Jev for routing, risk checks, and tool review.

The useful takeaway is straightforward: Jev’s first two weeks gave it the broadest early reach in Vercel AI Gateway history, according to Vercel. Developers evaluating it should focus less on the adoption record and more on whether its structured decisions improve a specific workflow.

The best choice depends on how often you use Gemini and which tools matter to you. For occasional questions, writing help, and light research, Plus may be enough. If you regularly work with complex research, generate media, or use AI across Google's apps, Pro can offer more room to work. Plan benefits and prices vary by country, so check Google's current offer for your account before subscribing.

Gemini Plus vs Pro at a glance

Google describes its access multipliers relative to subscribers who do not have an AI plan. They are not a promise of a fixed number of prompts per day. Individual tools have their own limits, and those limits can change. The US prices above are a snapshot, not a global price list. Check the official Google AI plans page for the current price and benefits in your region.

For a broader breakdown of every tier, including Ultra, see Saganote's Gemini pricing guide. This article focuses on the choice between Plus and Pro.

What does Google AI Plus include?

Google AI Plus adds more AI access than a standard Google account and bundles that access with 400 GB of storage for Gmail, Google Drive, and Google Photos. It also includes more access to selected Gemini features and creative tools, along with AI features in supported Google products. The exact set of tools depends on availability, language, account type, and region.

Plus makes the most sense if you want a paid Gemini plan but do not regularly hit its limits. It can suit everyday writing, brainstorming, summaries, study help, and occasional image or video creation. If you mainly need a chatbot for a few tasks each week, paying for Pro may not deliver enough extra value to justify the higher bill.

There is also a free option. Google has changed the capabilities available in its free Gemini tier over time, so compare the current free offering before upgrading. Saganote's guide to free AI tools in 2026 can help you decide whether a paid plan is necessary at all.

What extra value do you get with Google AI Pro?

Pro includes the benefits of Plus with expanded access to Google's advanced models and tools. Google's current plan information lists expanded access to Gemini 4 Argon for complex workflows, more access to Gemini 3.1 Pro and Deep Research, and additional capabilities across supported Google products. Pro also includes 5 TB of storage, compared with Plus's 400 GB.

That difference matters most when you use AI for work that takes several steps or requires repeated iterations. A larger usage allowance can reduce how often you have to wait, switch tools, or simplify a task because you reached a limit. It does not guarantee that every answer will be better, and Pro's model and feature availability still depends on Google's terms and rollout.

If access to Google's newest model is a key reason for upgrading, read Saganote's coverage of Gemini 4 Argon in Search and the Gemini app. Check the model's availability and limits in the subscription details rather than assuming every plan gets identical access.

Price difference: is Pro worth four times as much?

At the listed US monthly prices, Google AI Pro costs $15 more per month than Plus. That is $180 more over 12 months if both plans stay at those prices. Pro is also listed at four times the monthly price of Plus, so the extra features need to matter to justify the difference.

The storage gap is substantial: Pro offers 5 TB, while Plus offers 400 GB. If you already pay for storage or regularly run out of space for photos, video, and files, that bundle may account for part of Pro's value. If you do not need the storage, judge the upgrade mainly on AI access and the tools you actually use.

The wider market can help put that price in context. Saganote's AI pricing comparison looks across major providers, while ChatGPT vs Gemini vs Claude compares the assistants themselves. Those comparisons are useful if you are deciding whether to spend more on Gemini or pay for a different service.

Which plan should different users choose?

Choose Google AI Plus if...

• You use Gemini several times a week but rarely hit usage limits.
• You mainly need writing, summarization, brainstorming, and everyday help.
• You want extra Google storage without paying for the higher Pro tier.
• You want to try Google's paid AI features before committing to a more expensive plan.

Choose Google AI Pro if...

• You rely on Gemini for work or study most days.
• You use advanced models, Deep Research, or other demanding features often.
• You create media and want higher limits across supported creative tools.
• You need substantially more Google storage and will use the bundled 5 TB.

Stay on the free plan if...

• You only ask occasional questions or need basic writing assistance.
• You are not reaching the free tier's current limits.
• You do not need the extra storage or paid-only features.

What to check before subscribing

• Local price: Google's price and included storage can vary by country and offer.
• Feature availability: Some features are limited by region, language, age, or account type.
• Usage limits: A plan's access multiplier does not mean unlimited use. Check the specific limits for the tools you need.
• Storage needs: Consider how much space Gmail, Drive, and Photos currently use.
• Your existing subscriptions: Check whether you already pay for Google storage or another AI assistant.

If the free version is enough for your current needs, there is no reason to upgrade just because a paid plan exists. Google changes model access and plan benefits over time, so revisit the official comparison before renewing or switching tiers.

The verdict: Plus is enough for many people, Pro is for heavier use

Google AI Plus is the sensible starting point for people who want more Gemini access and extra storage at a lower price. Google AI Pro is easier to justify when its higher limits, advanced tools, and 5 TB of storage solve a real need. Compare your actual usage with the plan limits, then pay for the tier that removes a restriction you regularly encounter.

YouTube GIF comments are beginning to appear for some users, letting viewers respond to videos and Shorts with animated images instead of text alone. The rollout is gradual, so some people can see GIFs in comment threads without having the option to post one themselves.

How YouTube GIF comments work

The feature uses a GIF library inside the YouTube app. YouTube announced the change during its September Made on YouTube event, alongside other updates to viewer and creator features. The company has not clearly confirmed whether users will be able to upload their own animated files, so the supported option is currently the in-app library.

YouTube’s official Made on YouTube announcement introduced GIFs in comments as part of its wider set of community features. The YouTube Data API revision history also documents an image URL field for comments and says the feature will be available to all users in November 2026.

The rollout is still limited

The GIF option is not available to every account yet, and current reporting indicates that access is not restricted to Premium subscribers. The gradual rollout means availability can differ between users. If the option is missing, that does not necessarily mean the account is ineligible.

GIF comments add another way for viewers to react in YouTube’s comment threads. The change also fits a broader push to expand how people interact with video: YouTube has been developing conversational AI editing for Shorts and YouTube Create, custom video feeds built with Gemini, and A/B testing for three video cuts in YouTube Studio. It has also adjusted how it recommends re-uploaded Shorts that add no original value.

Other platforms are adding richer ways to respond and work with video, including TikTok comments with voice messages, polls, and photo carousels and an X video editor with multilingual captions and green screen. Related video workflows include timestamped comments in Google Drive, Google Photos Video Remix, and Instagram Edits’ AI assistant for creator ideas. For a separate Google Photos privacy update, see the Redact tool for blurring sensitive information on Android.

For now, YouTube GIF comments remain a gradual rollout rather than a feature every viewer can use. The company’s developer documentation points to November 2026 for availability to all users.

What Are Claude Code Mods?

Claude Code Mods are extensions that change how Claude Code behaves. Developers can use them to adjust what happens when Claude Code receives a prompt, calls a tool, asks for permission, or displays information.

The important distinction is that a Mod can do more than add a command or provide extra instructions. It can change how Claude Code handles an action while a session is running.

What Can a Mod Actually Do?

The best way to understand Mods is to look at practical examples. These are possible uses, not guarantees that every Mod will include all of these features.

1. Change prompts before Claude sees them

A Mod can rewrite a prompt before it reaches the model. A team might use this to add project-specific context or apply a consistent instruction format.

For example, a team could add a reminder to follow its coding standards whenever a task involves a particular repository. The Mod can apply that rule automatically instead of relying on every developer to remember it.

2. Block or change tool calls

A Mod can inspect a tool call and block it, rewrite it, or retry it. This can help teams apply workflow rules around commands and other actions.

Imagine a project where commands that affect production settings need extra checks. A Mod could detect those commands and require an additional step before allowing the workflow to continue.

This should not replace a careful permission setup. The behavior depends on how the Mod is written and configured.

3. Redact sensitive output

A Mod can remove secrets from tool output before Claude reads it. For example, it could look for API keys or tokens and replace them with a placeholder.

That can reduce the chance of sensitive values entering the model's context. It is not a complete security guarantee. Detection rules can miss secrets, and the Mod itself must be trustworthy.

4. Customize the interface

Mods can change parts of the interface that Claude Code displays. They can add buttons and input fields, or replace certain interface elements.

A team could create a panel that shows build status beside the coding conversation. Another Mod could provide a custom view for a workflow that developers use often.

5. Replace selected built-in features

Some Claude Code features are themselves implemented as Mods. Anthropic has said that the /diff feature is one example. Users can disable it through the plugin interface or replace it with a custom version.

This approach could make Claude Code more modular over time. Developers may be able to keep the parts they need and replace selected features with alternatives.

How Do Claude Code Mods Work?

Claude Code emits events as it works. Events can relate to tool calls, permission requests, prompts, or interface rendering.

A Mod registers functions for the events it wants to handle. When a matching event occurs, Claude Code runs the relevant function. Depending on the event and the Mod, that function can pass the original result through, change it, block it, or replace the normal behavior.

Several Mods can respond to the same event. Their order matters because one Mod may receive the event before another. That makes testing important, especially when multiple Mods are installed together.

Mods vs. Plugins, Hooks, and Skills

These terms are related, but they describe different parts of the Claude Code ecosystem.

• Mods: TypeScript functions that can change how Claude Code handles events and interface behavior.
• Plugins: Packages used to distribute extensions. A plugin can include Mods and other capabilities, depending on its contents.
• Hooks: Automation points that run logic at defined stages, such as before or after a tool is used. Traditional hooks are useful for repeatable checks, but they do not offer the same ability to rewrite events or replace interface features.
• Skills: Reusable instructions and workflows that guide Claude Code through a task. A skill can shape how Claude approaches work, but it is not the same as code that intercepts a tool call.

The simplest rule is to choose the least complex option that solves the problem. Use a skill when the main need is reusable instructions. Use a hook for a defined automation step. Consider a Mod when you need deeper control over event handling or the interface. Use a plugin to package and distribute extensions.

For broader context, see our guide to Claude Code's built-in side agent and our comparison of Kimi Code and Claude Code.

Examples of Useful Mods for Developers

Mods make the most sense when they solve a repeated problem. Here are a few examples of where they could help.

• Production safeguards: flag or block commands that target production resources.
• Secret filtering: redact likely credentials from tool output before it enters the model context.
• Build status panel: display CI/CD results beside the coding conversation.
• Team workflow rules: add checks that apply consistently across a project.
• Custom interface controls: add buttons or inputs for a workflow that developers use frequently.
• Alternative feature behavior: replace a built-in feature with a version that better fits a team's process.

A Mod should have a clear purpose. Avoid installing several extensions that change the same event unless you understand how they interact. Overlapping behavior can make a workflow harder to debug.

Are Claude Code Mods Safe?

Mods deserve careful review because they run with the same access to your machine as Claude Code. Anthropic says Mods are not sandboxed. A malicious or poorly written Mod could therefore create risks similar to running other untrusted code.

Treat a Mod like a program you are choosing to run, not like a harmless setting.

Before installing a Mod

• Check the source. Prefer Mods from authors and repositories you trust.
• Read the code when possible. Look for unexpected network activity, file access, or changes to permissions.
• Understand its purpose. Avoid installing a Mod that does more than its description requires.
• Test it in a low-risk project. Do not start with a production repository or sensitive credentials.
• Review how it handles permissions. A Mod should not quietly weaken safeguards you rely on.
• Keep it updated. Recheck the code and behavior when the Mod changes.

Anthropic also documents a built-in Mod called sec-default for Team and Enterprise plans and for machines using managed settings. It loads early to restrict certain risky behavior from installed Mods, including attempts to override permission-deny rules. Administrators can configure their own Mods to load first, but Anthropic advises keeping sec-default in the list if those restrictions are needed.

These controls help, but they do not make every third-party Mod safe. Teams should still review extensions and use plugin controls to limit which sources developers can install.

How to Get Started With Claude Code Mods

You do not need to build a Mod to benefit from the feature. You can install a plugin that includes one through the Claude directory or the /plugin interface in Claude Code.

1. Start with a clear need. Identify a repeated task or behavior you want to change.
2. Look for an existing plugin. Check its source, documentation, and permissions before installing it.
3. Install through Claude Code. Use the plugin interface or the official instructions for the plugin.
4. Test the behavior. Confirm that it works as expected in a non-critical project.
5. Review the result. Check that the Mod does not interfere with permissions, other Mods, or your existing workflow.

If you want to build your own Mod, start with Anthropic's official Claude Code Mods announcement and the official Mods source and examples on GitHub. The GitHub repository provides implementation details that are useful when you are ready to inspect the code.

For a related explanation of how AI tools connect to external services, read What Is Model Context Protocol (MCP)?. MCP and Mods solve different problems, but both are part of the wider ecosystem around extending AI-assisted development.

When Should You Use a Mod?

A Mod is worth considering when ordinary settings, instructions, or hooks cannot provide the control you need. It is especially useful when you need to inspect or change an event, alter part of the interface, or replace a built-in feature.

It may be unnecessary if you only need to add project guidance or run a simple check after a file changes. In those cases, a skill or traditional hook may be easier to maintain.

For teams, the decision should include maintenance and security. Someone needs to own the Mod, review updates, test conflicts, and document what it changes. A small extension can become a critical part of the development workflow if everyone depends on it.

The Bottom Line

Claude Code Mods give developers a deeper way to customize an AI coding workflow. They can change prompts, control tool calls, redact sensitive output, and reshape parts of the interface. They also provide a way to replace selected built-in features.

That flexibility comes with responsibility. Mods are executable code, not just configuration. Start with a specific need, choose trusted sources, test carefully, and keep security controls in place.

Google has launched Universal Cart in the United States across Google Search and the Gemini app. The feature lets shoppers save products from different retailers in one place, monitor price and stock changes, find relevant promo codes, and get payment-method suggestions based on eligible card benefits.

What Google Universal Cart can do

Universal Cart keeps shopping progress available across supported Google experiences. For example, a shopper can save one product while browsing AI Mode in Search and add other items through Gemini, then return to the same cart later.

• Save products across retailers: Keep items from different stores together instead of relying on separate tabs or screenshots.
• Track price and stock changes: Receive notifications when a saved product drops in price or becomes available again.
• Find possible savings: View relevant promo codes and suggestions based on payment-card benefits supported through Google Wallet.
• Choose how to check out: Use Google Pay with participating merchants or transfer items to a retailer's website to complete the purchase.

Google says the feature draws on its Shopping Graph, which contains more than 60 billion continually refreshed product listings. Participating checkout merchants include Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, and merchants using Shopify. Availability and checkout options can vary by merchant.

Checkout still involves retailers

Universal Cart does not replace the retailer. Customers can complete eligible purchases through Google Pay or continue on the merchant's own website. Retailers remain the seller of record and handle order confirmations and customer support.

The launch follows Google's broader push to add actions to its AI products. Google has also been expanding Gemini features in Chrome and testing ways for Gemini in Google Maps to help users place food orders. Those features show how Google is connecting AI-assisted discovery with tasks users may want to complete.

Read more about Gemini Chrome's Auto Browse rollout in the US and Google Maps' Gemini food-ordering feature.

Universal Cart is rolling out in the US. Google has not announced availability in other countries in its launch post.

The pitch is more than a cheaper design canvas. FigCraft combines local-first editing with tools that can be controlled through a command-line interface, a JSON control channel, and a Model Context Protocol (MCP) server. That combination puts it at the intersection of open-source creative software and the growing effort to let AI agents operate applications directly.

A Figma-style editor that keeps files local

FigCraft supports common interface-design building blocks, including frames, auto layout, components, variants, styles, variables, vector editing, and prototype connections. Its developer handoff view also offers measurements, annotations, and generated code for formats including CSS, HTML, React, SwiftUI, and Compose, according to the project's documentation.

The key distinction is how the app handles work. FigCraft stores designs as .figcraft archives on the user's disk and says it does not require an account or send telemetry. The app can work offline, and the project offers import and export for formats such as SVG, PNG, JPG, WebP, and PDF. Opening Figma .fig files is supported on a best-effort basis, not as a guarantee of perfect compatibility.

This local-first approach may appeal to independent designers, developers, and teams that want more control over where working files live. It also connects to a wider shift in creative software. Saganote has covered ArtCraft's broader set of open-source Adobe alternatives, a family of tools built around a similar idea: native creative applications developed in Rust rather than subscription-first cloud services.

The more unusual feature is control by AI agents

FigCraft says its tools and menus are exposed through more than 400 commands, which can be used through its CLI, JSON control channel, or MCP server. In practical terms, compatible agents could invoke application actions instead of only describing a design or generating code outside the editor. Readers new to the concept can learn what an AI agent is and how agentic software works.

That does not automatically make FigCraft an autonomous designer. Agent access is an interface for controlling the app; the quality of the result still depends on the agent, the instructions, and the design task. The approach does, however, fit a broader pattern in which AI tools are moving from chat windows into the software where work happens. Saganote has also examined Claude Cowork's move onto web and mobile and Claude Code's built-in browser for testing sites and navigating designs.

Figma itself is also connecting design and implementation more closely. Its Code Layers feature at Config 2026 brings executable code into the design canvas, a different approach to reducing the distance between a mockup and a working interface. FigCraft's agent-facing controls raise a related question: will future design tools be judged as much by how well software agents can operate them as by how comfortable their canvases feel to human designers?

Offline does not mean collaboration is solved

Local files are attractive for privacy and offline work, but they do not remove every collaboration problem. The FigCraft project describes live sessions that can be hosted from the app or on a server controlled by the user. Its current documentation also says it does not yet have Figma-style accounts, teams, or a cloud service. That is a meaningful limitation for organizations that depend on shared libraries, centralized permissions, review workflows, and managed file access.

The project's use of Rust is another part of its identity. Rust is a systems programming language often used where developers want direct control over performance and memory safety. FigCraft says it uses a native Rust engine rather than Electron or a web view. For background, Saganote's explainer on WebAssembly and why developers care about it covers another technology that can help software run across different environments, including browsers.

ArtCraft lists desktop builds for macOS, Windows, and Linux, along with a web build. The web version is based on WebAssembly, but availability across platforms should not be confused with feature parity or production readiness.

The project's own roadmap is the reality check

FigCraft labels itself alpha software and estimates that it is about 49% of the way toward replacing Figma for real work. The project says the core editor works end to end, while warning that .fig import is best effort and that its compatibility has not been checked against real Figma documents. That admission matters more than a long list of feature checkboxes: opening a design file is only useful if components, typography, assets, and layout behavior survive the move.

The same caution applies to the wider AI software wave. GPT-6's interactive apps in ChatGPT and the evolving capabilities of assistants such as ChatGPT, Gemini, and Claude show how quickly the boundary between asking software for help and interacting with software is changing. But a new workflow still has to prove that it is reliable enough for real projects.

Other recent stories show why independence from hosted services can be a product choice as much as a technical one. Saganote has reported on Meta's push to sell AI compute through cloud infrastructure and the removal of an Instagram AI photo feature after its opt-out design drew criticism. Those developments are not direct comparisons with FigCraft, but they underline how hosting, control, and defaults shape the way people use software.

A credible alternative, but not a drop-in replacement yet

FigCraft offers a clear proposition: a free, open-source design editor that can work offline, keep files local, and expose its tools to AI agents. That makes it worth watching for designers and developers who value local control or want to experiment with agent-driven workflows.

The harder test is whether the project can close the gap between a promising alpha and the dependable collaboration environment professional teams expect. Until file compatibility, collaboration, and day-to-day stability are proven on real projects, FigCraft is best treated as an early alternative to evaluate, not a guaranteed replacement for Figma.

The project, its current downloads, and its documented limitations are available on the official FigCraft page.

Grok Voice 3.0 customer support is a useful way to judge xAI's new voice model because support calls require more than natural-sounding speech. An agent must understand a caller through noisy phone audio, ask relevant follow-up questions, retrieve account information, and complete a task without losing the thread. xAI says Grok Voice 3.0 is built for that kind of real-time work, and that it already powers customer conversations at Starlink and voice interactions in Tesla vehicles.

The model launched on October 14, 2026, through xAI's Voice Agent API. The company says it ranks first on Artificial Analysis's Speech-to-Speech Index and costs $0.08 per minute, the same rate as Grok Voice 2.0. Those are meaningful claims, but the practical test is whether an automated agent can solve a customer's problem reliably, not simply produce a fast or convincing voice.

What Grok Voice 3.0 is designed to do

Grok Voice 3.0 is a speech-to-speech model. Rather than requiring an application to treat every conversation as a separate speech-recognition, text-model, and speech-generation job, it is designed to process spoken input and respond with speech in a real-time session. Developers can connect the agent to tools and business systems so it can do more than answer general questions.

That distinction matters in customer support. A basic voice bot might explain a return policy or read out troubleshooting steps. A connected agent could, if the business has granted the right permissions and built the integration, look up an order, check an account, issue an eligible credit, or start a replacement shipment. The model supplies conversation and reasoning capabilities; the surrounding software determines which real actions it is allowed to take.

What xAI says Grok Voice 3.0 can do at Starlink

The most concrete example in xAI's announcement is Starlink's phone support operation. In its Grok Voice 3.0 announcement, xAI says Grok Voice handles around 20,000 customer calls a day across multiple languages, resolves roughly 70% without a human agent, and converts about one in four sales calls before the caller hangs up. The company describes an agent that can verify customers, troubleshoot hardware, issue credits, and arrange replacement equipment. xAI has also published a separate Starlink case study describing how the system handles support and inbound sales calls.

These are company-reported operating figures, not an independent audit of every call. They nevertheless show the kind of workflow xAI wants customers to imagine: the voice agent does not merely explain a process, but carries a request through several steps until the customer has a result.

• Account verification: establish which customer is calling before exposing or changing account information.
• Troubleshooting: ask questions, interpret the answers, and guide the caller through relevant checks.
• Service recovery: apply an approved credit or arrange a replacement when the case meets company rules.
• Sales: answer product questions and help complete an order during the call.

The underlying challenge is orchestration. A successful call may require the agent to understand what the customer means, choose the next step, call the right tool, interpret the tool's response, and explain the result clearly. If any link in that chain fails, a natural voice can make the failure sound smoother without actually solving the problem.

For more context on the product category, see Saganote's explainer on how voice AI models work. It explains the broader technology behind systems that listen, reason, and speak during a live conversation.

Do the benchmarks prove it is better?

xAI cites Artificial Analysis testing dated October 13, 2026. In that evaluation, Grok Voice 3.0 scored 86.9 on the Speech-to-Speech Index, compared with 81.3 for Grok Voice 2.0, 83.2 for GPT-Live-1 (Astra, Medium), and 82.6 for Gemini 3.8 Live (Extended Thinking, High). The index combines several measures, including speech reasoning, task success, and human preference.

The results are encouraging, but they answer different questions. A high task-completion score does not guarantee that an agent will correctly handle every refund, a frustrated customer, or an unusual hardware problem. A conversational-dynamics score measures aspects of turn-taking, not whether the company's answer was correct or fair. And a benchmark's test set cannot represent every language, accent, telephone connection, or business workflow.

The most relevant figure for support operators may be task success, alongside the τ-Voice result for calls with accents, noise, and interruptions. Even then, companies should test the model on their own anonymized cases and measure successful resolution, repeat calls, escalation quality, customer satisfaction, and mistakes that require remediation.

Why response speed matters, but cannot be the only goal

xAI says Grok Voice 3.0 has latency below one second and reports 40% lower latency than GPT-Live-1 on the cited τ-Voice comparison. In a phone conversation, latency is the delay between a person speaking and the agent responding. Long pauses can make a system feel broken, cause callers to repeat themselves, or make it harder to interrupt a mistaken answer.

But the fastest response is not always the best response. A billing dispute or a request to change an account may require the agent to verify information or check a policy before answering. For those cases, a short pause followed by a correct answer is better than an instant guess. Support teams should optimize for useful resolution and safe action, not speed in isolation.

What does Grok Voice 3.0 cost?

xAI lists Grok Voice 3.0 at $0.08 per minute of audio, or $4.80 per hour, through the Voice Agent API. The company says the rate is unchanged from Grok Voice 2.0. At that listed model rate, the raw audio charge for 10,000 minutes would be $800 before other applicable costs.

These are simple estimates based on the published per-minute rate, not a full call-center budget. A real deployment may also need phone service, software integration, monitoring, secure data access, human escalation, and ongoing testing. Before comparing providers, check whether each price includes telephony and the surrounding agent platform, or only the model's audio processing.

For background on xAI's previous voice release, read Grok Voice Think Fast 2 and its focus on faster real-time agents. The new release keeps the same published per-minute price while updating the model behind the API.

How the rest of xAI's voice stack fits together

Grok Voice 3.0 is one part of a larger set of audio tools. A company may need a live conversational agent, transcription of recorded calls, generated speech for an app, or a way to configure an agent without building every component from scratch. These products overlap, but they are not interchangeable.

• Live voice conversations: Grok Voice 3.0 is aimed at speech-to-speech sessions where the agent listens and responds in real time.
• Transcription: Grok Voice Transcribe 2.0 is the relevant Saganote coverage for xAI's speech-to-text model, which turns audio into text rather than handling the entire conversation itself.
• Building an agent: xAI's broader agent products include Grok Bot, its cloud-based AI teammate, and later coverage of Grok Bot's expanded access. Those are broader agent workflows, not simply voice-model releases.
• Voice technology fundamentals: the voice AI models explainer provides background for readers comparing speech recognition, speech generation, and live speech-to-speech systems.

For readers following the broader product ecosystem, Saganote has also covered Grok 4.7's launch and context window, Grok 4.6's pricing and long-running agents, and Grok 4.6 on Google Cloud Vertex AI. These releases concern general-purpose models and deployment options, while Grok Voice 3.0 focuses specifically on live audio interactions.

Another related development is Grok Build's wider availability for eligible plans. It is a developer workflow product rather than a customer-support voice model, but it illustrates how xAI is expanding its software tools around the Grok family.

For a broader comparison of personal-agent products, see Meta Muse vs. Grok Bot vs. OpenAI Dots. That comparison is useful for distinguishing assistants that perform ongoing tasks from a voice agent designed to handle a live phone conversation.

Where AI customer support can still go wrong

A high-performing model can still make support worse if a company gives it too much authority or measures the wrong outcome. Common risks include incorrect account actions, misunderstood requests, weak handling of edge cases, and conversations that keep customers trapped in automation when they need a person.

• Incorrect actions: tool calls should have permission checks, confirmation steps, and clear limits on refunds, account changes, and shipments.
• Poor escalation: callers need a clear route to a human when the issue is sensitive, ambiguous, or unresolved.
• Privacy and identity: the system should verify callers appropriately and avoid exposing personal information to someone who is not authorized to receive it.
• Uneven performance: real calls contain accents, background noise, interruptions, and network problems that may not be fully represented in benchmark tests.
• Misleading success metrics: a call that ends quickly is not necessarily a resolved call. Teams should track repeat contacts, complaints, reversals, and customer feedback as well.

The best deployment is likely to start with well-defined, repetitive workflows, while leaving high-impact or uncertain decisions to trained staff. Human escalation should be part of the design, not a fallback added after customers start reporting problems.

Who should consider Grok Voice 3.0?

Grok Voice 3.0 is worth evaluating for teams that handle a high volume of predictable phone requests and can connect the agent to reliable internal tools. Examples include order status, basic troubleshooting, appointment changes, product questions, and selected sales workflows. Its listed per-minute price makes it possible to estimate model costs before a pilot, though total cost depends on the surrounding system.

It is a less obvious fit for organizations that cannot safely expose the necessary systems to an automated agent, have complex cases that routinely require human judgment, or lack the staff to audit call outcomes. Those teams may get more value from transcription and agent-assist tools that help human representatives rather than replace the conversation.

The practical next step is a controlled pilot. Select a narrow set of call types, compare the agent against existing outcomes, review failures manually, and expand only when the system meets clear standards for accuracy, escalation, and customer experience.

Frequently Asked Questions

The bottom line

Grok Voice 3.0 is notable because xAI is positioning it as a voice agent that can complete support workflows, not just speak naturally. Its benchmark results, reported Starlink use, and predictable per-minute price make it a serious option for companies exploring automated phone support. The evidence is promising, but the deciding factor for any business will be its own measured resolution rate, error rate, escalation quality, and customer experience.

The update addresses scams that begin with ordinary conversation and only reveal their intent over time. Google says Enhanced protection will begin rolling out in the coming weeks; it is not available to every user immediately. It also fits into Google’s broader September Android feature drop, which added other system-level improvements.

How Gemini helps identify complex scams

With Enhanced protection enabled, Google Messages can use Google's cloud Gemini models to evaluate the context of conversations with unknown senders. Instead of relying only on familiar scam patterns, the additional analysis is designed to help identify suspicious intent that may emerge gradually across multiple messages.

Google says this cloud processing runs through Private AI Compute, which it describes as an encrypted, secure environment that prevents access to the processed data, including by Google. That is Google's stated privacy model; the announcement does not provide an independent assessment of the new feature. The approach also illustrates how the company is bringing Gemini agents into everyday business workflows, though this Messages feature is focused on scam detection rather than task automation.

Link checks and warnings get an upgrade

Enhanced protection also adds two changes for links sent by unfamiliar numbers:

• Gemini Nano link checks: The on-device model uses real-time web signals to assess unfamiliar links before they are opened, including sites that may not yet appear on known malicious-site lists.
• Clearer link warnings: Redesigned alerts provide more context and an extra moment to pause before opening links from numbers that are not saved in contacts.

These features build on existing protections in Google Messages. Google's wider Android scam-protection updates also include expanding scam alerts in chat notifications to more Android manufacturers, safeguards for sensitive information during screen sharing, and protection against apps silently enabling call forwarding. The link-scanning approach is relevant to other tools, too, including Truecaller Scam Checker, which focuses on identifying risky links, numbers, and messages.

Who gets Enhanced protection?

Google says Enhanced protection is an opt-in setting that will roll out over the coming weeks. The link-checking features are described for Pixel 10 and newer, Samsung Galaxy S26 series, Galaxy Z Flip 8 and Z Fold 8, and other select devices. Availability may vary by device, and Google has not said that every Android phone will receive every part of the feature.

The update arrives as Google continues expanding its messaging app across Android devices. It comes as Samsung Messages is being discontinued, making Google's app more central to messaging on Galaxy phones. Google is also expanding AI features across its products, from Gemini Spark in Chrome to the Gemini app for Windows. Messaging remains a key platform battleground, as the broader iOS 27 update also brings changes to Apple’s Messages app.

The key change is the move from detecting suspicious message patterns on a device to optionally using cloud AI to interpret a longer conversation. That may help with scams designed to build trust slowly, but the feature's real-world effectiveness will depend on how reliably it catches malicious conversations without flagging legitimate ones. Google’s broader AI ecosystem also includes paid plans detailed in its Gemini pricing guide, although the announcement does not say Enhanced protection requires a paid Gemini subscription. Google has also been improving how Android devices move between platforms, including the iPhone-to-Android transfer process.

Google says its existing Scam Detection features already flag more than five million suspected scam calls and texts each month. That figure comes from Google, and the company has not yet shared a separate performance measure for Enhanced protection.

Gemini 4 Argon is now available to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. Google announced the wider release on October 13, expanding the model beyond its earlier limited access. Developers and enterprise customers can also use Argon through the Gemini API, AI Studio, and Google's business platforms.

Where Gemini 4 Argon is available

For everyday users, the main change is that Argon can be selected in the Gemini app and used through Search's AI Mode, provided the account has an eligible Google AI Pro or Ultra subscription. The model is not included in the free Gemini tier. Google says Argon is designed for complex, multi-step work involving coding, finance, legal research, cybersecurity, and science.

Developers can access the model using the ID gemini-4-argon through the Gemini API and Google AI Studio. Google's developer documentation lists a context window of up to 1,048,576 tokens and says long outputs can reach up to 1,048,576 tokens when long-decode mode is used. The default synchronous maximum output is lower, so the million-token figure should not be read as the standard output limit for every request.

A wider rollout, but not free access

The launch expands the number of places where people can use Argon, but access still depends on the product and subscription. Google's official announcement says AI Pro and Ultra subscribers can use it in the Gemini app and Search AI Mode, while developers and enterprises have separate API and platform access.

Google is positioning Argon as a model for difficult, longer-running tasks rather than just quick chat. Its claims about benchmark performance come from Google's own evaluations, so they should be treated as company-reported results rather than independent proof that Argon is best for every task.

This release follows Google's broader Gemini model rollout, including three new Gemini models and an early tease of Gemini 4. Saganote has also covered the model's one-million-token output capability and coding benchmarks and access status.

The subscription distinction matters because Google has also changed what users can access for free. Its Gemini free-tier limits and current Google AI Plus, Pro, and Ultra pricing help explain which features require a paid plan. Argon availability may also differ by product: Google's Gemini API documentation lists API access and technical limits separately from consumer-app availability.

Google's expanding product lineup includes the NotebookLM-to-Gemini Notebook transition and its plans for Gemini models in more workflows. Previous coverage has also examined delays around Gemini 3.5 Pro and Google's reported Gemini 3.5 Pro launch timeline. Those stories provide context for how Google is bringing newer models to users, but they do not change Argon's current eligibility requirements.

For users choosing a plan specifically for Argon, the key point is simple: Google has expanded access to its Gemini app and Search AI Mode, but free users should not expect the same model availability.

The steps below are based on Waze's current help documentation. Menu labels can vary slightly between Android and iPhone, and some features work differently when Waze is connected to a car display.

1. Navigate to a saved contact without typing an address

If a friend or family member's home address is saved in the phone's contacts, Waze can use that address as a destination. This is useful when arranging a pickup or visiting someone whose address is already stored in the contact card.

How to use it

1. Open Waze and swipe up to find Drive to friends and family.
2. Tap Next, then allow contact access when prompted.
3. Return to the map and tap the search bar.
4. Type the contact's name and select the matching address.
5. Choose Go now to start navigating, or Leave later if you want to plan the trip.

The important detail is that the contact needs a street address saved in the phone. A name and phone number alone are not enough for Waze to know where to route you. You can revoke contact access later in your phone's app-permission settings.

Official instructions: Drive to a phone contact in Waze.

2. Record navigation directions in your own voice

Waze can record custom voice prompts, so navigation can sound like you, a friend, or a family member instead of the default voice. The recording replaces supported spoken prompts; it does not mean the app will generate new speech in that person's voice.

How to record a voice

1. Open Waze on your phone.
2. Open the menu and go to Settings.
3. Select Voice and sound, then open Current voice.
4. Choose Add a voice or New voice, depending on the app version.
5. Follow the prompts and record each phrase.
6. Play the recordings back, save the voice, and give it a name.

You do not have to record every phrase at once. Start with the directions you use most, and make sure each recording is clear. Waze's help page explains how to share or delete a custom recording as well.

For Android Auto or CarPlay, set up the voice on the phone first. Waze says to disconnect from the car system while recording or changing the voice.

Official instructions: Record your own voice directions.

3. Schedule a drive before the day gets busy

Waze can plan a trip around the time you need to arrive, rather than only calculating a route for right now. It estimates when you should leave using expected traffic and can send a reminder when departure time approaches.

How to plan a drive

1. Open Waze and tap the menu.
2. Choose Plan a drive.
3. Tap the + button and search for the destination.
4. Select the date and the time you want to arrive.
5. Save the planned drive and make sure Time to leave notifications are enabled.

For example, if an appointment starts at 9:00 a.m., set the arrival time rather than guessing how early to leave. Waze can adjust the suggested departure based on traffic conditions. Check the planned trip before leaving because traffic, weather, and road conditions can still change.

You can also plan a drive after checking a route by selecting Leave later. Waze's support documentation notes that planning a drive is not supported directly on Android Automotive, and some steps require the phone app.

Official instructions: Plan a drive in Waze.

4. Connect your calendar to get departure reminders

If appointments already live in your phone's calendar, you can connect the calendar to Waze instead of creating every trip from scratch. Waze can show upcoming events in planned drives and send a time-to-leave notification when it is time to go.

How to connect a calendar

1. Open Waze on your phone.
2. Tap the menu and select Plan a drive.
3. Tap Connect calendar.
4. Allow calendar access, then confirm.
5. Check that upcoming events include a complete destination address.

The address matters. Waze says an event needs a full location, including street name, house number, and city, for a departure reminder to work. If Waze shows Set address beside an event, review the event's location and correct it.

Official instructions: Sync your calendar with Waze.

5. Make Waze quieter, or troubleshoot missing voice audio

Waze does not have to announce every turn in the same way for every driver. Its sound settings can reduce spoken guidance, leave alerts on, or mute audio. This is useful when frequent instructions become distracting, or when a driver wants to hear hazard alerts without full turn-by-turn narration.

How to change the sound mode

1. Open Waze and start or preview a route.
2. Tap the sound icon.
3. Choose Less chatty if it is available and you want only essential guidance and alerts.
4. Choose Alerts only if you want alerts but not spoken turn-by-turn directions.
5. To change the volume, open Settings, then Voice and sound, and adjust Guidance volume.

The available choices can depend on the device and selected voice. If Less chatty does not appear, check the current voice and the sound options shown in the app rather than assuming the feature is available in every configuration.

Waze also documents a search command, ##@cctts, for clearing its text-to-speech cache when voice audio is not working properly. This is a troubleshooting step, not a general performance booster. Try the standard sound settings first, and use the command only if you are troubleshooting spoken directions.

Official instructions: Set your sound preferences.

Which Waze feature should you try first?

Start with the feature that removes the most friction from your routine. Contact navigation is useful for frequent visits, custom voice directions make navigation more personal, and planned drives are helpful for appointments with a fixed arrival time. Calendar sync is best for people who manage their schedule digitally, while sound controls help tailor how much guidance the app gives.

For a broader look at how Waze fits into everyday navigation, see Saganote's guide to Waze vs. Google Maps and their different navigation styles. Drivers interested in recent Waze changes can also read about Waze's motorcycle mode and personalized navigation updates.

What to know before relying on these features

• Keep Waze updated because menu names and feature availability can change.
• Review contact and calendar permissions, and share only the access you want to grant.
• Enter complete addresses in contacts and calendar events.
• Treat departure reminders as estimates, not guarantees. Check traffic before leaving.
• Set up custom voices and calendar access while parked, not while driving.

These features are most useful when they are configured before a trip. A few minutes spent saving accurate addresses, setting a preferred voice, and planning an arrival time can make Waze easier to use without changing the way you normally navigate.

Facebook is introducing Facebook Your Algorithm, a set of controls that lets people explain their content preferences in their own words. Announced by Meta on October 13, the feature is designed to help users influence what appears in Feed and Reels rather than relying only on likes, comments, watch history, and other engagement signals. Meta says recommendations can update in real time, but the rollout is still underway.

How Your Algorithm works

The dedicated Your algorithm area is accessible from the three-dot menu on Facebook posts. People can add and manage topics they want to see in Feed and Reels. Meta is also expanding feedback options: after someone chooses “Interested” or “Not interested,” Facebook may ask follow-up questions so they can explain what they liked or want to see less often.

Meta says large language models help its ranking systems interpret the context and intent behind these written preferences. The company says changes can affect recommendations within the same session. It has not published independent evidence showing how much the controls improve recommendation quality, and the announcement does not mean every account has access yet.

A broader shift toward user-controlled recommendations

The feature follows a wider move toward letting people shape recommendation feeds. YouTube, for example, has been testing more personalized viewing experiences, including custom feeds powered by Gemini. Spotify has also given eligible U.S. Premium listeners more ways to adjust their Taste Profile, while Google has expanded personalization across Search, Discover, and News.

Social platforms still make most recommendations through automated ranking. X has also had to address how its algorithm ranked replies from mutuals, while YouTube has changed how it handles re-uploaded Shorts that add little original value. These examples show that recommendations affect not only what people watch, but also which creators and posts get visibility.

For creators, the practical effect remains uncertain. People may be able to steer their feeds more directly, but Meta has not quantified how much a written preference will outweigh other ranking signals. Its rollout announcement also does not specify when the feature will reach every user.

The change fits into Meta’s wider creator and editing tools, including Facebook Creator Studio on iPhone, AI suggestions in Instagram Edits, and automatic trimming for Instagram Reels drafts. Meta has also previously removed a photo feature after users reacted to its opt-out default, a reminder that control settings matter as much as new features.

Other platforms are experimenting with related discovery and publishing changes, from replacing music on existing Instagram posts to changes in how video platforms rank content. Facebook Your Algorithm’s key test will be whether written feedback produces a noticeable, lasting change without requiring people to constantly correct their feeds.

Meta’s announcement is available in its official newsroom post.

The X Subscription bundles Grok, Cursor, and X Premium benefits into one account, giving users a single plan for AI chat, coding tools, and social features. The company says its $30-per-month Premium tier includes the benefits of SuperGrok, Cursor Pro, and X Premium+, which previously cost $90 per month combined. The company outlined the offer in its official announcement and on the X account announcement.

How the X Subscription works

The new lineup starts with a free plan and adds Plus for $8 per month, Premium for $30, Super for $100, and Ultra for $200. Paid tiers combine different levels of AI usage with X features such as a checkmark, creator tools, and fewer ads. The full details are on the official X Subscription pricing page.

Grok and Cursor draw from one shared weekly usage pool. Subscribers can use more of that allowance for AI conversations or coding, depending on their needs. The allowance also covers Grok Bot on eligible tiers. However, third-party models in Cursor, including Claude and GPT models, are billed separately through usage top-ups rather than included in the shared pool.

Existing subscribers do not have to switch immediately

Existing SuperGrok, Cursor, and paid X subscribers can keep their current plans or switch through account billing settings when the option appears. Eligible customers who switch receive a prorated refund for unused time on the old plan. There are exceptions for subscriptions billed through app stores, so customers should check the official switching instructions before changing plans.

The bundle is most useful for people who already pay for several of these services. Readers comparing individual costs can also review Saganote’s guides to Cursor pricing and Grok pricing. For someone who only uses one product, the main question is whether the bundled features and shared usage allowance justify the monthly cost.

This explainer separates Meta’s published claims from independently reported results and outlines how to compare Muse Wonder with other coding models. It is not an independent hands-on test: the results below describe available documentation and published benchmark claims, not experiments run by Saganote.

What Is Meta Muse Wonder?

Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026. The models are built for personal and professional AI-agent work, including coding, document analysis, and tasks that require several steps rather than a single response. Meta says the models use Active Memory to carry useful context between conversations and Active Focus to keep working toward a goal, try alternate approaches, recover from errors, and ask for clarification when needed.

For developers, the distinction between a chatbot and an agent matters. A chatbot may suggest code or explain a bug. An agent can also use tools, inspect a repository, edit multiple files, run tests, and continue through several stages of a task. The quality of that process depends on the model, the tools it can access, the permissions it receives, and the tests used to check its work.

Meta’s official Muse Wonder announcement describes the models’ intended uses and safety approach. Its developer documentation provides the current model variants, supported inputs, and access details.

Muse Wonder vs. Muse Wonder Full

These details reflect Meta’s documentation checked on October 13, 2026, and may change. A one-million-token context window lets a model accept a large amount of text and project context in one request, but it does not guarantee that the model will use every detail correctly. Context capacity and reliable reasoning are different measures.

Meta lists Muse Wonder Full for the Power Usage and Maximum Usage Muse Code plans. At the time of checking, those plans were listed at $50 and $200 per month respectively. Subscription prices, usage limits, and eligibility should be checked on the official Muse Wonder page before committing to a plan.

What Do the Benchmarks Show?

Meta’s model page publishes results across several categories, including coding, long-horizon software engineering, computer use, memory, professional tool use, and specialist work in areas such as finance and law. The breadth of those tests reflects the product’s intended role: Muse Wonder is meant to handle workflows that stretch beyond generating a short code snippet.

Meta’s published comparison includes models from Anthropic and OpenAI. Those scores are useful as a starting point, but they need context. Benchmarks can differ in task selection, model configuration, tool access, inference settings, and scoring rules. A score from one evaluation does not automatically predict performance on a team’s codebase or preferred development workflow.

The distinction between company-reported and third-party results also matters. Meta’s model page identifies independent third-party runs for Muse Wonder Full on DeepSWE and SWE-Marathon. Other entries in the model comparison should be read according to the evaluation notes and report linked from the official benchmark page, rather than treated as independent confirmation by default.

A benchmark is most useful when it matches the task being considered. A terminal-coding benchmark may reveal how a model handles software tasks in a controlled environment. It does not, by itself, prove that the model will be the fastest, cheapest, or most dependable assistant for every developer.

How Muse Wonder Should Be Tested Against Other Models

A fair comparison should use the same tasks, repository snapshots, tool permissions, and acceptance tests for each model. The exact model versions and test dates should be recorded because AI products change quickly. Avoid comparing a model with browsing and terminal access against another model that has been denied those tools.

• Bug fixing: give each model the same reproducible bug and check whether the proposed fix passes existing tests without creating regressions.
• Multi-file implementation: ask each model to add a feature across several files, then run the same test suite and inspect the diff.
• Unfamiliar repository: test whether the model can locate relevant code, explain its plan, and make a targeted change without unrelated edits.
• Long-running task: measure whether it can recover from tool errors, maintain requirements, and finish the assignment.
• Context retention: provide requirements across multiple interactions and check whether the model remembers constraints accurately.
• Cost and speed: record elapsed time, token usage, retries, human corrections, and the cost of a successful result.

Run each task more than once when practical. Report completion rate and common failure modes, not only the best result. A useful coding comparison should also include the amount of human intervention required: a model that produces plausible code quickly may still be a poor choice if developers must spend substantial time debugging it.

Which Models Belong in the Comparison?

Claude coding models and OpenAI’s current coding-capable models are natural candidates because Meta itself compares Muse Wonder with models from those labs. Google’s Gemini can also be included if the selected version supports a comparable setup. The final shortlist should be chosen immediately before testing, with model names, versions, access tiers, and tool configurations documented.

For readers comparing coding tools rather than models alone, Saganote has related coverage of Meta Muse Code and its approach to large codebases, Claude Sonnet 5.5 code review, and Claude Sonnet 5 vs. Opus 4.8 benchmarks and pricing. Those comparisons address adjacent questions, but their results should not be treated as a substitute for testing Muse Wonder under the same conditions.

Developers evaluating Meta’s API can also read about Meta’s developer API and Muse Spark 1.1. Muse Wonder and Muse Spark serve different workload priorities: Meta recommends Wonder for longer, more complex tasks where capability matters more than token cost, and Spark for cost-sensitive or latency-sensitive workloads. Check current documentation before choosing, since model availability and pricing can change.

Access, Pricing, and Practical Limits

As of October 13, 2026, Meta lists Muse Wonder on the Meta Model API and in Muse Code. The API price is $4 per million input tokens and $20 per million output tokens, with cached input priced at $0.20 per million tokens on the model page. Muse Wonder Full is available through eligible Muse Code subscriptions rather than as a standalone API model.

Token rates are only one part of the total cost. Long-running agents may make many model calls, use tools, retry failed approaches, or generate large outputs. A practical evaluation should estimate the total cost of completing a task, not multiply the advertised rate by a single prompt and assume that represents the whole workflow.

Access also affects reproducibility. Record whether each model was tested through an API or a coding agent, which tools were enabled, and what subscription or usage tier was used. If Muse Wonder Full is available only through a subscription while another competitor is accessed through an API, make that difference clear rather than presenting the setup as identical.

Who Should Consider Muse Wonder?

Muse Wonder is worth evaluating for developers who regularly work across large repositories, need agents to complete multi-step tasks, or want to automate engineering workflows with tool use. Its large context window and agent-oriented design are relevant when a task requires coordinating many files or preserving requirements across a long session.

Teams with short, repetitive, latency-sensitive requests may not need the more capable model for every task. A lower-cost model can be a better fit when the work is simple and the output is easy to verify. The right choice depends on the quality of completed work, failure recovery, total cost, speed, security requirements, and how well the model fits existing tools.

For broader agent workflows beyond coding, readers can explore Saganote’s coverage of Meta Muse as a personal AI agent. An agent’s ability to act across applications introduces additional questions about permissions, privacy, and approval steps that coding benchmarks alone cannot answer.

What to Watch Next

The most useful next evidence will be reproducible evaluations from developers using real repositories, clearer methodology for the published benchmark results, and practical reports about reliability over long tasks. Meta says additional safety evaluations will be available in a separate report; those results should be reviewed alongside capability claims before teams use agents for sensitive workflows.

Until comparable hands-on results are available, the safest conclusion is limited: Muse Wonder is designed for complex coding and long-running agent tasks, and Meta’s published benchmarks position it as a competitor to leading models. Those claims provide a reason to test it, not proof that it will outperform alternatives on every project.