Gemini 3.5 Pro Is Live: 2M Context, Deep Think in Vertex Preview

Gemini 3.5 Pro launched on July 17, 2026 in a limited Vertex AI preview — matching its widely reported target date after a third delay that had pushed the release from June. Multiple independent trackers confirm the model went live July 17, with a 2.1-million-token context window and Deep Think reasoning mode. Google has not published a standalone launch blog post or confirmed public API pricing as of July 18.

What Actually Happened on July 17

Gemini 3.5 Pro went live on July 17 in a limited Vertex AI preview — confirming the widely reported date after a chaotic six-week build cycle that included at least three separate delay announcements. The launch was quiet by Google standards: no standalone blog.google announcement, no public model card, and no official API pricing page as of this writing. Independent model trackers confirmed the July 17 release, with a 2.1-million-token context window placing it second among all tracked frontier models for context depth.

The limited preview status means general availability is still pending. Enterprise customers with Vertex access can test the model now; broader developer access through Google AI Studio and the public Gemini API has not been announced. This pattern matches how Google rolled out Gemini 3.1 Pro — limited Vertex access first, broader API release weeks later.

Confirmed: 2M Context Window and Deep Think

The two headline specs from the pre-launch leaks — the 2-million-token context window and the Deep Think reasoning layer — have been confirmed by independent model trackers, though Google has not yet published an official model card with verified benchmark scores. The context window of 2.1 million tokens (2,097,152 to be exact) is the second-largest of any tracked frontier model. For practical reference, 2 million tokens can hold an entire software codebase, a year of enterprise email, or tens of thousands of rows of structured data — all in a single prompt without chunking.

Deep Think, the step-by-step reasoning mode, is confirmed present. Access tier and pricing remain unconfirmed — the pre-launch leak placed it on a $250/month Ultra plan; the actual launch tier has not been officially stated. Until Google publishes a pricing page, treat any tier assignment as provisional. What is clear is that Deep Think operates similarly to reasoning modes from OpenAI and Anthropic: trading response speed for deeper problem-solving on multi-step tasks like mathematical proofs, long-horizon code refactors, and policy analysis.

Kimi K3 Launched the Same Day — What It Means for the Competition

Moonshot AI launched Kimi K3 on July 16–17, the same window as Gemini 3.5 Pro, and the juxtaposition is striking. Kimi K3 is a 2.8-trillion-parameter open-weight model — the world’s first open 3T-class model — with a 1-million-token context window and native vision. It landed at $3.00 per million input tokens and $15.00 per million output tokens, undercutting Claude Opus 4.8 ($5/$25) and GPT-5.5 ($5/$30) on both sides.

In blind Arena evaluations, developers chose Kimi K3 ahead of all U.S. frontier models on front-end coding tasks, and independent rankings place it fourth overall — trailing only Claude Fable 5, GPT-5.6 Sol, and closely edging past Claude Opus 4.8. Full open weights are scheduled to drop by July 27, 2026, according to Fortune’s July 17 coverage.

The comparison with Gemini 3.5 Pro is instructive: Kimi K3 wins on context depth by 1 million tokens (Gemini’s 2.1M vs Kimi’s 1M), Gemini wins on transparency (Kimi K3 is open-weight; Gemini 3.5 Pro is proprietary), and on price Kimi K3 is drastically cheaper if Google’s unconfirmed leaked pricing of ~$12–15/$36–45 per million tokens holds. If you’re a developer choosing between them today, Kimi K3 is available via public API right now; Gemini 3.5 Pro is not. If you’re an enterprise user with Vertex access, Gemini 3.5 Pro’s tighter Google ecosystem integration is the differentiator to test.

What the Gemini 3.5 Pro Leaks Actually Said

As reported by TechTimes on July 13, Gemini 3.5 Pro targeted July 17 after a full architectural rebuild. The headline pre-launch leak was the 2-million-token context window, now confirmed by trackers. For comparison, Anthropic’s Claude Fable 5 ships with a 200K standard window, and OpenAI has not publicly confirmed the context size for GPT-5.6 Sol.

The second major leak concerned Deep Think, a reasoning layer reportedly reserved for the $250-per-month Ultra plan. Now that the model is live, the spec appears confirmed in function — the pricing tier has not been officially stated. Reports also suggested the rebuild specifically targeted three weaknesses: mathematical reasoning, SVG scene generation, and image quality.

On API pricing, a leaked post circulating on X suggested ~$12–15 per million input tokens and ~$36–45 per million output tokens. These figures remain unconfirmed — Google has still published no official pricing page as of July 22. For context, Gemini 3.1 Pro currently costs $2/$12 per million tokens; the leaked Gemini 3.5 Pro pricing would represent a 6–8× premium, reflecting the compute overhead of Deep Think reasoning.

Why Google Rebuilt the Model From Scratch

Gemini 3.5 Pro was originally slated for late June 2026 — making the July 17 launch more than six weeks late. According to StartupFortune, Google delayed after scrapping its base model entirely, citing “quality refinements after early enterprise testing.” A full rebuild this late in a release cycle is unusual — it signals either serious performance problems in the original model or unusually high ambitions for the replacement. Geeky Gadgets reported on July 15 that the model had encountered a third delay, with hallucinations and inconsistent outputs as specific concerns. Despite that, the July 17 limited preview appears to have shipped.

The Talent Losses That Raised the Stakes

The delay arrived alongside the most significant talent exodus Google has seen in its AI division. On June 18, Noam Shazeer — Gemini co-lead and co-author of the original Transformer architecture paper — left for OpenAI. The following day, John Jumper, the 2024 Nobel laureate who led AlphaFold, announced his move to Anthropic, along with fellow researchers Jonas Adler and Alexander Pritzel.

The exits triggered a roughly 7% drop in Alphabet’s stock on June 22, erasing an estimated $270 billion in market value across two trading sessions — a stark signal of how tightly investor confidence in Google is now linked to its AI leadership pipeline.

The departures matter beyond the stock chart. Shazeer’s institutional knowledge of transformer architecture runs deep, and Jumper’s move to Anthropic directly strengthens the team building Claude. That both men left during the very period when Gemini 3.5 Pro was being rebuilt from scratch adds context the spec sheets won’t include. The limited-preview launch on July 17 is Google’s first public signal that the rebuild survived those losses.

Google Ships Three Flash Models on July 21 — Gemini 3.5 Pro Still Absent

Google released three Flash-tier models on July 21, 2026 — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber — while Gemini 3.5 Pro remained unreleased five weeks past its July 17 limited preview. All three sit in Google’s Flash tier, tuned for speed, cost, and high-volume agentic work rather than maximum reasoning depth.

Gemini 3.6 Flash is the primary workhorse update. Priced at $1.50 per million input tokens and $7.50 per million output tokens, it costs less than 3.5 Flash while generating 17% fewer output tokens on the Artificial Analysis Index — meaning lower total cost per agentic task. On coding benchmarks, 3.6 Flash scores 49% on DeepSWE versus 37% for 3.5 Flash, and improves ML research performance (MLE Bench: 63.9% vs. 49.7%). Computer use is now a built-in client-side tool via the Gemini API.

Gemini 3.5 Flash-Lite targets high-throughput agentic workloads. At $0.30 per million input tokens and $2.50 per million output tokens, it is Google’s cheapest 3.5-class model, running at 350 output tokens per second per Artificial Analysis. On coding benchmarks it outperforms Gemini 3 Flash — 54.2% vs. 49.6% on SWE-Bench Pro and 74.0% vs. 65.1% on OSWorld-Verified — making it a faster and cheaper option for workloads previously routed to Gemini 3 Flash.

Gemini 3.5 Flash Cyber, a fine-tuned variant of 3.5 Flash optimized for finding and fixing cybersecurity vulnerabilities, will be available exclusively to governments and trusted partners via Google’s CodeMender platform as part of a limited-access pilot. Google did not publish general API pricing for Cyber; the model is not available to individual developers.

On Gemini 3.5 Pro, Google’s official blog stated the model is “currently testing with partners” and the company plans to “make it broadly available as soon as it’s ready.” Bloomberg reported meeting internal performance targets has been a challenge — consistent with the model’s history of missed deadlines since its original late-June target.

Google also announced it has begun “the most ambitious pre-training run yet” for Gemini 4 — the first public acknowledgment of the next-generation model’s training timeline.

What We Still Don’t Know

Pricing remains the biggest open question. Google has published no official API pricing page for Gemini 3.5 Pro as of July 22. The leaked $12–15/$36–45 per million token figure may or may not hold when general availability is announced. General availability itself has no announced date. No independent benchmark scores have been verified with a source URL — third-party ranking sites report the model is live but show no confirmed leaderboard position. We will update this article when Google publishes an official model card, pricing page, or GA announcement.

The timing follows a crowded frontier-model summer: OpenAI’s GPT-5.6 Sol, Terra, and Luna family entered limited preview on June 26, Moonshot AI’s Kimi K3 launched July 16–17, and Google’s own Gemini 3.5 Flash has been serving as the family’s speed tier since earlier this year.

FAQ

Updated: August 5, 2026