What you need to know
- Launch day is here; the launch is not — yet. As of this morning IST there is no official announcement, model card or API listing for Gemini 3.5 Pro. The 17 July date itself is third-party reporting, never confirmed by Google.
- The official record is two sentences long. At I/O on 19 May, Google shipped Gemini 3.5 Flash and said Pro was "being used internally" and coming "next month". That June window slipped.
- Everything else is leak — the 2M-token context window, the Deep Think reasoning layer, autonomous workflow capabilities, the $250-per-month Ultra tier gating and every benchmark number you have seen.
- The practical move is the same either way: build on Flash now, keep routing abstracted, and have your own evals ready to run the hour the model card publishes.
The official record is thin — and that is the story
Strip away the speculation and Google's on-the-record position on Gemini 3.5 Pro fits in a paragraph. At Google I/O on 19 May 2026, the company announced the Gemini 3.5 series and made 3.5 Flash generally available the same day — in the Gemini app, AI Mode in Search, the Gemini API and its Antigravity development platform. On Pro, the official announcement said only: "We're also hard at work on 3.5 Pro. It's already being used internally, and we look forward to rolling it out next month."
"Next month" meant June. June came and went. Trade coverage — including Startup Fortune and TechTimes — subsequently reported that the launch had moved to 17 July, and Google has at no point confirmed that date. We checked again this morning: Google's Cloud documentation for the Gemini Enterprise Agent Platform lists Gemini 3.5 Flash and the older Gemini 3.1 Pro, and no 3.5 Pro entry. No model card, no pricing page, no gemini-3.5-pro model ID anywhere in the public API documentation.
One caveat on timing before anyone declares the date missed: Google launches on US Pacific time. A Mountain View morning announcement lands around 9.30pm IST and 5pm in London, so "no launch as of this morning" and "launched on 17 July" can both end up true. What will not change in the intervening hours is the gap this article is about — the distance between what Google has said and what the leak economy has claimed on its behalf.
Confirmed vs leaked: the ledger
This is the state of every major claim as of publication. "Confirmed" means it appears in an official Google announcement or documentation; "leaked / reported" means it traces back to unnamed sources via third-party outlets, however many of them repeat it.
| Claim | Status | Where it comes from |
|---|---|---|
| Gemini 3.5 Pro exists and is in internal use | Confirmed | Google's I/O announcement, 19 May 2026 |
| Gemini 3.5 Flash generally available (app, Search, API) | Confirmed | Google's I/O announcement, 19 May 2026 |
| Series positioned for agents, coding, long-horizon tasks | Confirmed | Official Gemini 3.5 series framing |
| 17 July launch date | Leaked / reported | TechTimes, Startup Fortune, HackerNoon; unnamed sources |
| 2-million-token context window | Leaked / reported | Multiple outlets; absent from all official docs |
| Deep Think reasoning layer on a $250/month Ultra tier | Leaked / reported | Single-track leak, widely recycled |
| Autonomous workflow / agentic task decomposition | Leaked / reported | Unnamed sources via trade press |
| Original base model scrapped, pretraining restarted | Leaked / reported | Trade reporting; Google has not commented |
| Limited enterprise preview on Vertex AI since late June | Leaked / reported | Industry trackers, per AIToolsReview |
| API pricing | Unknown | Nothing published; all figures circulating are estimates |
| Benchmark scores | Unknown | Google published no Pro benchmarks at I/O or since |
Benchmark screenshots for Gemini 3.5 Pro are already circulating on X and LinkedIn. None of them can be traced to an official Google publication — Google has published no Pro scores at all. If a chart cannot be traced to an official model card, do not let it into a procurement deck or a migration decision.
Why the rebuild story matters more than the specs
The most consequential of the unconfirmed claims is not the context window — it is the reporting that Google DeepMind scrapped the original Gemini 3.5 Pro base model and restarted portions of pretraining. According to BigGo Finance and the TechTimes reporting above, engineers opted for a ground-up rebuild after finding structural weaknesses in areas such as recursive tool-calling and complex scene generation, rather than shipping an incremental update. Google has neither confirmed nor denied any of this, so treat the specifics as unverified — but the delay itself is a matter of public record: Google telegraphed a June arrival at I/O, and June passed without a release.
If the rebuild reporting is even directionally right, it tells you two useful things. First, Google is optimising this generation for agentic reliability — the failure modes reportedly being fixed are exactly the ones that break long-horizon agent loops. That is consistent with the confirmed positioning of the 3.5 series as models "for agents and coding". Second, a restarted pretraining run under launch pressure raises the odds of a staged rollout: a Gemini-app and AI-Studio release first, with API general availability, full pricing and rate limits trailing by days or weeks. Plan for that shape rather than a single clean GA moment.
The competitive backdrop explains the pressure. This is the most crowded fortnight of the year for frontier releases — TechTimes notes that GPT-5.6 went public on 9 July and xAI opened Grok 4.5 the same day — so every additional day of slippage costs Google mindshare among exactly the builders it wants back. When we analysed the wilder end of the rumour mill last month in our look at the Gemini 3.5 Pro Omni and world-model speculation, the honest conclusion was that the claims outran the evidence. Launch week has not changed that; it has only raised the stakes on the evidence arriving.
What builders in Bengaluru and London should actually do
The temptation on launch day is to refresh the changelog and plan architecture around leaked numbers. The better play, whether you are a two-person team in HSR Layout or a platform group in Shoreditch, is to position yourself so the model card — whenever it lands — is a one-day integration rather than a re-architecture.
- Build against Gemini 3.5 Flash now. Flash is confirmed, generally available and, per Google's announcement, outperforms the previous 3.1 Pro flagship on coding and agentic benchmarks. Anything you ship on Flash today inherits Pro cleanly later — same API surface, same tooling. We covered the model and its Spark Agent tooling in our I/O report on Gemini 3.5 Flash.
- Keep model routing abstracted. If swapping
gemini-3.5-flashforgemini-3.5-prorequires more than a config change, fix that this week. Teams that route by task class — cheap model for extraction, frontier model for planning — absorb new releases in hours, not sprints. - Prepare your evaluation suite before GA. Fifty representative tasks from your own workload, scored automatically, beat any public leaderboard. Run them the day the model appears and you will know by evening whether the upgrade is real for you — while your competitors are still arguing about headline benchmarks.
- Budget against confirmed prices only. Nothing official exists for Pro pricing; circulating figures are estimates anchored to Gemini 3.1 Pro's published rates. Do not sign customer contracts, set unit economics or promise margins on leaked numbers — in either rupees or pounds.
- If Deep Think ships, treat reasoning budget as a dial, not a default. The leaked reasoning layer would follow the same economics as every other thinking-mode model: cost scales with deliberation. Our guide to prompting reasoning models and setting thinking budgets applies directly on day one.
One regional note worth flagging for both markets. Frontier launches rarely arrive everywhere at once: consumer app features, AI Studio access and Vertex AI availability have historically landed on different schedules in different countries, and premium tiers sometimes reach the US weeks before India or the UK. If the leaked Ultra-tier gating for Deep Think proves accurate, check availability and local pricing in your own region before you promise a capability to a client in Mumbai or Manchester — "launched" in a headline does not always mean "available on your account today".
Write your "Pro migration" pull request today against Flash: config flag, eval harness wired in, rollback path tested. When the model ID goes live you merge one line, run the evals and decide on data. That is the entire launch-day playbook, and it works in any timezone.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The bottom line
Gemini 3.5 Pro's launch day arrived with the launch itself still unverified and the specification sheet still written entirely by leakers. That is not a reason to dismiss the model — Google's own framing of the 3.5 generation, and the calibre of the confirmed Flash release, suggest Pro will be a serious frontier contender the moment it ships. It is a reason to hold two lists in your head: the short one Google has signed its name to, and the long one it has not. Build against the first, prepare for the second, and let the model card — not the rumour mill — make your engineering decisions. We will update our coverage the moment official documentation lands.