Gemini 3.5 Pro Targets July 17 - and Google Rebuilt It From Scratch to Get There

Google DeepMind is targeting July 17 for Gemini 3.5 Pro, a model the company chose to rebuild entirely from scratch rather than ship what it had. The same week, DeepSeek's legacy API aliases - deepseek-chat and deepseek-reasoner - stop responding on July 24. Any code still calling those endpoints after that date returns errors, with no extension announced. Two things happening in the same week, one of them optional to track and one of them not.

Why Google Scrapped the Base Model

Gemini 3.5 Pro was supposed to ship in June. At Google I/O on May 19, Sundar Pichai told the audience to "give us until next month." It didn't arrive. The delay landed in the same two weeks as Noam Shazeer's departure to OpenAI and John Jumper's move to Anthropic - two of Google DeepMind's most senior researchers - which knocked about $225 billion off Alphabet's market cap in a single session.

The bigger surprise was what caused the delay. Google DeepMind reportedly abandoned the Gemini 2.5 Pro base model entirely and ran a full pre-training cycle from scratch. Three gaps drove the decision: mathematical reasoning, SVG scene generation, and image quality. Incremental fine-tuning hit a ceiling on all three. Running a new pre-training run at frontier scale costs hundreds of millions of dollars and takes months of compute time - Google made that call anyway, which says something about how far short the prior candidate fell.

What 3.5 Pro Is Supposed to Deliver

Reported specs (from third-party sources, not official Google documentation): a 2 million token context window, double the 1 million cap on Gemini 2.5 Pro; a Deep Think Reasoning Layer for multi-step logic; and autonomous workflow capabilities for chaining complex coding tasks. As of July 8, the public Gemini API lists only gemini-3.5-flash and gemini-3.1-pro-preview. No model card, no benchmark, no pricing has been published. July 17 is a widely reported target date, not a signed launch post.

The context window claim is the one to watch most closely. At 2 million tokens, the model could process roughly 1.5 million words in a single prompt - a full large codebase, a year of meeting transcripts. The engineering challenge is that transformer attention scales quadratically with sequence length, and independent research has documented that model accuracy degrades for information buried in the middle of very long contexts even when it technically fits. Until independent evaluators publish long-context retrieval benchmarks, 2 million tokens is a headline, not a guarantee. Context window size and actual retrieval accuracy are different things - a lesson that's shown up in evaluations of every model that's claimed a record.

DeepSeek's July 24 Deadline Is Not Optional

DeepSeek deprecated its old model aliases when V4 launched in April. The grace period ends July 24, 2026, at 15:59 UTC - after that, any API call using deepseek-chat or deepseek-reasoner returns an error. No extension has been announced.

Migrating is a one-line change: update the model parameter to deepseek-v4-pro or deepseek-v4-flash on the same base URL with the same API key. One catch worth flagging: deepseek-reasoner maps to V4-Flash (thinking mode), not V4-Pro. A team that used deepseek-reasoner for heavy reasoning tasks and assumes the rename preserves capability parity will end up on Flash-tier reasoning at Flash prices. Anyone running complex agentic workloads on deepseek-reasoner should explicitly evaluate V4-Pro before migrating, not just swap the alias.

For teams self-hosting, none of this applies - V4 is MIT-licensed and the weights are available. For teams on the hosted API, July 24 is a hard cutover with nothing else to track on the roadmap. GPT-5.6 landed last week and Gemini 3.5 Pro is inbound - it's a dense two weeks for anyone evaluating their AI stack, and the DeepSeek deadline is the one that doesn't wait for the reviews to come in.