AI coding tools split into three layers: assistants that suggest code, agents that finish tasks, and the models that power both. Saganote checked prices on November 21, 2026. Vendor pages and Saganote's own older guides disagree on several numbers, and this guide flags each conflict where it appears in the text.
Assistants, Agents, and Models: How the Category Splits
An assistant completes or edits code inside an editor. An agent pursues a goal by choosing steps, using approved tools, and checking results, and it keeps going until the task is done or a person has to decide.
Autocomplete is the oldest layer. Copilot's completions and next edit suggestions do not use credits, and Cursor's pricing guide says unlimited Tab on paid plans does not mean unlimited model usage, since frontier models drain the allowance faster.
Models sit underneath. A model is the engine, and the same Claude, GPT, Gemini, Grok, or Kimi model can power several products, which is why rankings of coding models and tool comparisons answer different questions.
Agentic work has a definition problem too. Agentic AI covers systems that act toward goals instead of answering one prompt, though such systems still make mistakes and need supervision, and coding is where that supervision shows up as code review.
Context is the shared budget. A context window holds the system prompt, messages, tool definitions, tool results, and the reply, so long agent sessions fill it quickly, and a 1M-token window is not 1M tokens of reliable memory.
The Main Tools Compared
Overlap is large. Each vendor lists several surfaces, so the table records structure from vendor pages read on November 3 and November 21, not rankings.
| Tool | Home surface | Other surfaces named in vendor docs | Instruction file | Free tier |
|---|---|---|---|---|
| Claude Code | Terminal | VS Code and Cursor extension, JetBrains plugin, desktop app, web, Slack, GitHub Actions | CLAUDE.md, can read AGENTS.md | None, paid plans only |
| Codex | ChatGPT app and CLI | IDE extension, cloud tasks, GitHub, Slack, Linear | AGENTS.md | No free tier named on OpenAI's page |
| Cursor | AI code editor | CLI, cloud and background agents, web and mobile, Slack, GitHub, Linear | Rules | Limited free plan |
| GitHub Copilot | IDE | GitHub website, GitHub Mobile, CLI, desktop app, cloud agent | Custom instructions | Free plan with 2,000 completions a month |
A four-way comparison sorts the tools by where the agent works: terminal, editor, cloud, or pull request. Habit decides first. Pricing, controls, and review fit then break ties, and Codex documents the clearest approval modes, from read-only to a flag that removes the sandbox entirely.
Claude Code calls itself an agentic coding tool that reads a codebase, edits files, runs commands, and integrates with development tools, per its overview page. Surfaces include a terminal, VS Code and Cursor extensions, a JetBrains plugin, a desktop app that needs a paid subscription, and the web. Most surfaces need a subscription or Console account.
Codex is included in ChatGPT plans, per OpenAI's page, and it runs as Codex in ChatGPT, Codex cloud, an IDE extension, and a CLI. OpenAI's product map covers ChatGPT, GPT models, the Responses API, and Codex, and its Agents SDK serves developers who build their own agents. Codex has no free tier on that page.
A June comparison of Copilot, Cursor, and Tabnine reported 4.7 million paid Copilot subscribers in January 2026, up 75% year over year, and more than 1 million paying Cursor users. That comparison also lists Tabnine's agentic tier at $59 per user per month, with air-gapped deployment as its enterprise edge. SpaceX's $60 billion Cursor acquisition changed Cursor's ownership.
Cloud Agents and Team Channels
All four tools now run remote work. Codex cloud runs tasks in isolated containers managed by OpenAI, and each task keeps its own workspace. GitHub says Copilot's cloud agent can research a repository, plan changes, and implement them in the background using a cloud sandbox or a GitHub Actions runner, while Cursor lists background agents and Claude Code lists a web version and cloud scheduled tasks.
Network access is off by default. Codex's documentation says the agent runs with network access turned off by default, so a cloud task cannot reach package registries or internal services unless the environment allows it. Review stays human, since the agent returns a branch or pull request for approval.
Teams already work in chat tools and code hosts. Claude Code accepts an @Claude mention in Slack, Codex lists Slack and Linear integrations, and Cursor lists Slack, Linear, and GitHub. GitHub Copilot reached Microsoft Teams for coding tasks, and Slack Code runs coding agents in team channels.
Permissions and Enterprise Controls
Codex documents the clearest approval modes. An auto preset lets it read files, edit, and run commands in the working directory, and it asks before leaving the workspace or using the network. Stricter options include read-only and on-request, and a flag labeled not recommended removes the sandbox and approvals entirely.
GitHub says a branch protection rule can prevent Copilot from creating or updating pull requests, and a session without write access is read-only. Cloud sessions on Actions keep data 28 days. Teams that already govern work through GitHub rules inherit those controls without extra setup.
Codex lists a Compliance API, audit events, managed configuration, and roles for enterprises. Claude Code's overview page does not cover permission modes or data policies, and Cursor's agent pages were blocked from automated reading when Saganote checked on November 3. Direct checks with each vendor are the safer route.
Models Behind the Tools
BenchLM's coding composite, refreshed November 6, 2026, ranks Claude Opus 5.5 first at 82.8, with Claude Sonnet 5.5 at 81.9 for half the price. A ranking of coding models by published benchmarks and price per task lists 13 of the 157 models BenchLM tracks. Scores are BenchLM's own blend.
| Model | BenchLM coding score | Input and output price per million tokens | Note |
|---|---|---|---|
| Claude Opus 5.5 | 82.8 | $4 and $20 | Top of the composite |
| Claude Sonnet 5.5 | 81.9 | $2 and $10 | Half the price of Opus 5.5 |
| Gemini 4 Carbon | 81.3 | $2 and $10 | BenchLM marks the score as estimated |
| DeepSeek V4.1 Flash | 63.3 | $0.30 and $1.20 | Open weights |
| Kimi K3 | 58.9 | $3 and $15 | Open weights |
| Grok 4.6 | 58.6 | $2 and $6 | Cheapest closed model in that ranking |
Terminal-Bench 4.0 reorders the list, with Sonnet 5.5 at 70.6% and Opus 5.5 at 66.4%, and Anthropic reported both figures itself. Treat cost claims the same way.
Claude Haiku 5.5 costs $0.10 input and $0.50 output per million tokens up to 100K prompt tokens, per the October 8 report, and Anthropic positions it as a subagent beside Opus 5.5 and Sonnet 5.5. Check Anthropic's page before relying on it.
What Benchmarks Show and Do Not Show
Benchmarks measure models, not whole tools. SWE-bench Verified scores run from 88.6% to 97.0% on Vals AI's bash-only harness, which Morph treats as saturated, and BenchLM describes its coding composite as a weighted blend that leans on SWE-bench Pro and DeepSWE.
SWE-bench Pro numbers need extra care. BenchLM lists 89.9% for Opus 5.5, but Morph says that figure comes from llm-stats alone and that Anthropic does not report SWE-bench Pro, while Scale's standardized board topped out at 61.5% for Muse Spark 1.1 as of its newest entry on July 9.
Cost per task is the thinner evidence. Proximal's FrontierSWE v2 runs, reported through OrcaRouter, averaged $98.87 per run for Opus 5.5, $129.36 for Gemini 4 Argon over 10.6 hours, and $1,029.65 for GPT-6 Astra, though cache hit rates were low in a non-production setup. Anthropic's claim that Opus 5.5 matches GPT-6 Astra on Terminal-Bench 4.0 at about 40% of the cost is a vendor claim.
Models also swap inside tools. A walkthrough of using Kimi K3 in GitHub Copilot shows one tool running a different lab's model, which is another reason to compare tools and models separately.
Extending Tools With MCP and Skills
Anthropic published MCP in November 2024, and servers now give agents tools for repos, databases, browsers, and docs. Servers inherit the host app's permissions. Setup for GitHub, Playwright, Context7, and database servers differs across Claude Code, Cursor, and Codex, and read-only modes and narrow tokens are the shared advice.
Apple shipped a Safari MCP server on July 1 with 17 tools for AI agents that debug in Safari Technology Preview 247. Apple has not announced stable Safari timing.
Instruction files carry project rules between sessions. Claude Code can read an existing AGENTS.md file in place of CLAUDE.md, which eases using two tools on one repository, and both Claude Code and Codex support subagents, MCP servers, hooks, and reusable skills.
A skill is a folder with a SKILL.md file, kept in the .claude/skills/ directory for a project or the ~/.claude/skills/ directory for personal use. CLAUDE.md holds persistent facts, hooks run event-driven checks, and subagents take delegated work. Skills hold repeatable procedures.
What the Tools Cost
Four vendors use three billing models. Claude and ChatGPT sell subscriptions with usage windows, Copilot sells monthly credits worth one cent each, and Cursor sells weekly pools with third-party models billed on top at API list prices.
| Tool | Free | Entry paid plan | Higher tiers | Usage model |
|---|---|---|---|---|
| Claude Code | Not on Free | Claude Pro, $20 monthly or $17 billed annually | Max $100, Ultra $200 | Shared limits every 5 hours and weekly, plus usage credits |
| Codex | None named | ChatGPT Plus, $20 | Pro $100, Business $20 per user | Weekly usage set by plan |
| Cursor | Limited | Plus, $8, introductory | Premium $30, Super $100, Ultra $200 | Multiples of Plus, third-party models billed separately |
| GitHub Copilot | 2,000 completions, limited chat and agent | Pro, $10, with $15 in credits | Pro+ $39 with $70, Max $100 with $200 | 1 AI credit equals $0.01, completions use none |
Copilot's credits replaced premium requests on June 1, 2026, each worth one cent, and Pro includes 1,500 credits, Pro+ 7,000, and Max 20,000. Completions stay unmetered on paid plans. GitHub's first metered billing cycle drew developer anger over surprise bills, though that report's dollar figures conflict between its headline and body.
Cursor's pricing guide lists Pro at $20, Pro+ at $60, and Ultra at $200 as of September 2, while Cursor's own pricing page on November 21 shows Plus at $8, Premium at $30, Super at $100, and Ultra at $200. Cursor's page shows Plus as an introductory, Grok-only plan. Verify on cursor.com before buying.
Free tiers differ in kind. Copilot Free offers 2,000 completions a month with limited chat and agent use, and GitHub's plans page lists the Copilot CLI as included on Free. Verified students get Copilot Student at no cost. Cursor's free plan lists limited agent and chat requests, and Claude Code is not on Claude's Free plan.
Claude's page says Claude Code and chat draw from the same usage pool, and Max includes $100 a month in API credits while Ultra includes $200. OpenAI's page says Plus includes Codex usage for focused coding sessions each week, and Pro adds higher limits for full workdays. Neither vendor publishes limits as numbers.
Prices were checked on November 21, 2026 against vendor pages. Cursor's tiers differ between its own page and Saganote's September guide, Codex documentation had moved to a host that was not read, and Copilot Business and Enterprise seat prices came from a summary of GitHub's docs. Confirm on the vendor page before buying.
How to Choose
Match the tool to the work. A developer who lives in a terminal fits Claude Code or the Codex CLI, one who reviews diffs in an editor fits Cursor or Copilot, and a team that governs work through pull requests fits Copilot's GitHub rules.
Tools can share a window. Claude Code ships an extension for VS Code and for Cursor, so a second tool does not require a second editor, and developers comparing AI coding tools can start from the four-way table above.
Vibe coding means describing software in plain language and judging the result by running it, and its starter checklist covers version control from the first prompt, secrets kept out of prompts, least access for the agent, and approval before destructive steps. Review stays the human step.
Risk data comes mostly from vendors. A July 2025 METR trial found experienced open-source developers 19% slower with AI tools across 246 tasks while they believed they were 20% faster, and CodeRabbit's December 2025 analysis of 470 open-source pull requests found AI co-authored code had about 1.7 times more major issues. A Replit agent deleted a database despite instructions not to, per SaaStr's founder.
| Job | Fit per vendor docs | Note |
|---|---|---|
| Terminal-heavy scripting | Claude Code, Codex CLI | Both support subagents, MCP servers, hooks, and skills |
| Reviewing diffs in an editor | Cursor, GitHub Copilot | Editors keep review close |
| Governing work through pull requests | GitHub Copilot | Branch protection rules apply |
| Delegated background tasks | Codex cloud, Copilot cloud agent, Cursor background agents, Claude Code web | Review stays the human step |
| Requests from team chat | Claude Code in Slack, Codex in Slack and Linear, Copilot in Teams | Check each vendor's controls |
Every Guide in This Series
- Claude Code vs Codex vs Cursor vs GitHub Copilot: Which AI Coding Tool Fits Which Developer: working styles, surfaces, and a price snapshot.
- GitHub Copilot vs Cursor vs Tabnine: 2026 AI Coding Assistant Guide: three assistants compared on price, agents, and deployment.
- Best AI Models for Coding in 2026: Ranked by Published Benchmarks and Price per Task: the BenchLM ranking and cost evidence.
- Claude Haiku 5.5 Is Fast and Cheap. That Matters More Than Its Benchmarks: a small model for subagent work.
- Vibe Coding Explained: What It Is, the Tools People Use, and How to Start Safely: prompt-driven building and a safe-start checklist.
- How to Use Kimi K3 in GitHub Copilot: swapping a different lab's model into Copilot.
- Best MCP Servers for Developers: What Each One Does and How to Set It Up: server picks and config for three clients.
- Model Context Protocol: What MCP Is and Why Every AI Platform Now Supports It: the protocol behind the servers.
- Apple Ships a Safari MCP Server That Gives AI Agents Direct Browser Access for Debugging: Safari debugging with 17 tools.
- 5 Claude Code Skills That Give Developers Superpowers in Everyday Coding: skill folders and five examples.
- GitHub Copilot Pricing 2026: Plans and AI Credits: plans, credits, and metered features.
- GitHub Copilot's First Metered Bill Just Landed. Developers Are Furious.: the first billing cycle after credits.
- Cursor Pricing 2026: Hobby, Pro, Ultra, and What $20 Actually Buys: tiers, pools, and billing catches.
- What Is an AI Agent? A Simple Guide to AI That Can Act: the agent loop explained.
- What Is Agentic AI - And Why It's About to Change How You Work: a plain-language primer.
- What Is a Context Window: Why AI Forgets and How Tokens Work: tokens, limits, and caching.
- What Does OpenAI Do? A Complete Guide to ChatGPT, Models, APIs, Agents, and Codex: OpenAI's product map.
Prices, models, and ownership moved quickly in 2026, and AI coding tools will keep shifting after this November 21 check. One Copilot promotion ends November 30. GitHub's plans page advertised 30% off Copilot Fusion usage until that date, and the plan table above does not include the discount.



