Gemini 3.5 Pro has not shipped. July 17, 2026 the date the internet spent two weeks treating as confirmed has come and gone with no model card, no API listing, and no official benchmark from Google. As of this writing, Google's own spokespeople are describing the model as still being tested with partners, not released.
Who this is for: developers and technical leads who paused a model selection decision waiting on Gemini 3.5 Pro, and anyone trying to separate confirmed fact from leak-driven hype before they plan a Q3 stack.
Who this isn't for: if you're just looking for a chatbot to answer everyday questions, Gemini 3.5 Flash is already live and free to try you don't need to wait for anything in this article.
What's Actually Confirmed vs. What's Rumor
Almost every specific claim circulating about Gemini 3.5 Pro the 2 million token context window, Deep Think reasoning mode, and any benchmark numbers comes from third party reporting and unnamed sources, not from an official Google announcement. Google has confirmed the model exists and is used internally, and that's about the extent of the official record.
Here's the split between what Google has actually said and what's been leaked:
| Claim | Status |
|---|---|
| Gemini 3.5 Pro exists and is in development | Confirmed by Google |
| In limited Vertex AI enterprise preview | Confirmed by Google |
| 2M token context window | Reported, not confirmed |
| Deep Think reasoning mode | Referenced in I/O materials, no spec sheet |
| July 17 launch date | Rumor / leak, not an official date |
| Pricing ($12-15 input / $36-60 output per 1M tokens) | Community estimate only |
| Benchmark scores | Unpublished |
Treat every number you see attached to "Gemini 3.5 Pro" right now as a placeholder. None of it is from a Google model card.
The Timeline: A Flagship That Keeps Slipping
Google announced Gemini 3.5 Pro at I/O on May 19, 2026, positioning it as the replacement for the old Ultra tier the model meant to cover the hardest agentic and reasoning work. Sundar Pichai told developers on stage to "give us until next month." That was the June target. June came and went with no public release.
By late June, coverage described Pro as sitting in a limited Vertex AI enterprise preview available only to select customers, with Google citing quality refinements after early enterprise testing as the reason for the delay. The rationale reported wasn't cosmetic: early testers flagged real problems in advanced reasoning, coding, and long horizon task execution, along with token efficiency issues where the model burned through more tokens than expected to reach an answer.
The most striking detail is what Google reportedly did next. Rather than patch the existing Gemini 2.5 Pro based architecture, DeepMind scrapped that base model entirely and restarted pretraining on a native Gemini 3 foundation, specifically to close gaps in multi step mathematical reasoning, recursive tool calling stability, and SVG scene generation. Recursive tool call stability is not a minor detail here it's the exact requirement an agentic coding model has to nail, and it's the use case Google has staked the entire 3.5 generation on.
Why the Delay Actually Hurts More This Time
A June slip would have been forgivable in isolation. A July slip lands in a much worse competitive position, because the rest of the field didn't wait.
- GPT 5.6 (Sol, Terra, Luna tiers) launched publicly on July 9, 2026.
- Claude Fable 5 shipped on June 9, 2026, and is already on its second pricing structure.
- Grok 4.5 opened to the public on July 9, per Elon Musk's own announcement.
- DeepSeek's V4 family is targeting its own stable release and a mandatory API migration by July 24.
Three of the most watched model events of the year converged in the same two week window, and Google's flagship was absent from all of it. As of July 16, 9to5Google reported Google is "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners" language that describes an internal QA phase, not a company on the verge of a public launch announcement.
The One Thing Google Has Actually Shipped: Gemini 3.5 Flash
It's worth separating the delayed flagship from what's already live and working. Gemini 3.5 Flash launched at I/O on May 20, 2026, and has official, Google confirmed benchmark numbers: it outscored Gemini 3.1 Pro on Terminal Bench 2.1 (76.2% vs. 70.3%) and on MCP Atlas (83.6% vs. 78.2%), at a price of $1.50 per million input tokens and $9.00 per million output tokens.
That creates an awkward internal problem for Google, independent of anything competitors are doing. If the cheap, shipped Flash model is already winning agent benchmarks against last year's Pro tier, the unreleased 3.5 Pro has to justify a premium price with real reasoning depth and long-context quality not just a bigger number on a spec sheet nobody's seen yet.
Practical Takeaway: What to Actually Do Right Now
Don't build a Q3 plan around a model you can't test. Here's the sane way to handle the wait:
- If you need a Google model today: use Gemini 3.5 Flash for high volume agent loops, coding assistance, and everyday production work it's confirmed, cheap, and already outperforming last year's Pro tier on several benchmarks.
- If you need Pro tier reasoning today: Gemini 3.1 Pro Preview remains the closest real, testable Google option while 3.5 Pro is unavailable.
- If your workload needs the hardest long horizon agentic work now: this is exactly the gap Claude Fable 5 and GPT 5.6 Sol are positioned to fill, and both are live with published benchmarks you can test against your own tasks today.
- Don't lock in a decision based on leaked numbers. A 2M context window and a Deep Think mode mean nothing for your workload until Google publishes a model card you can actually run evals against.
- Set a re check date, not a wait and see mindset. Revisit this in two weeks. If Gemini 3.5 Pro still hasn't shipped by then, treat the "July 17" date the same way the market already treated "June": a target Google missed, not a deadline it kept.
For a full breakdown of how the models that have actually shipped stack up on pricing and benchmarks, see our Ultimate 2026 LLM Battle comparing DeepSeek V4 Pro, GPT 5.6 Sol, and Claude Fable 5.
No comments: