Google DeepMind finally has a date, and it is running low on runway to hit it. Gemini 3.5 Pro — the model Google reportedly tore down to the studs and rebuilt rather than ship as an incremental update — is targeting general availability on July 17. That is this coming Friday. As of this weekend, the public Gemini API still does not list it, and Google has not published a model card, API documentation, or a pricing page to make any of it official.
What is on the reported spec sheet is genuinely aggressive: a 2-million-token context window, roughly double the working memory of anything else at the frontier, plus a “Deep Think” extended-reasoning layer built for the hard, multi-step problems where today’s models still fold. The number that will decide who actually uses it, though, is not a benchmark. Deep Think — the mode that would make the model worth the hype — reportedly sits inside Google’s Ultra subscription, which runs as high as $250 a month.
The timing is the whole story. Gemini 3.5 Pro was unveiled at Google I/O in May with a June target; a full architectural rebuild dragged it into mid-July. It now arrives barely a week after GPT-5.6 went public and Grok 4.5 undercut the field on price. Google is walking into the most crowded week the frontier has ever seen — late, and with its best feature priced like enterprise software.
Our take: The 2-million-token window is the real weapon. It changes what fits in a single prompt — a whole codebase, a quarter of filings, an entire case file — without the retrieval scaffolding everyone bolts on today. But Google is making the same bet OpenAI and xAI just made: give the cheap tier away and charge serious money for the reasoning that actually closes the gap. At $250 a month, Deep Think is not competing with a chat subscription. It is competing with hiring an analyst. Until Google posts a model card, though, all of it is a target, not a shipped product — and this launch has slipped before.
What to watch
- Does July 17 hold? The launch has already moved once. A model card, API docs, or a live pricing page is the only thing that turns “targeting” into “shipped.”
- The real Deep Think price. “Up to $250 a month” is a ceiling, not a number. Whether there is a usable middle tier decides how many people ever touch the reasoning mode.
- Long-context accuracy. A 2-million-token window is a headline; the test is whether it stays sharp deep into the context, where long-context models have historically degraded.
- The floor, not just the ceiling. With Grok 4.5 dragging frontier capability down the cost curve, Google’s cheap tier may matter more than its expensive one.