What Gemini 3.5 Pro Actually Ships
Google's Gemini 3.5 Pro is now in limited enterprise preview on Vertex AI, announced at Google I/O in May and slipping past its June GA target as of June 23. Two capabilities define the model architecturally. The first is a 2-million-token context window — double Gemini 3.5 Flash and the largest deployed in any production frontier model to date. The second is Deep Think, a reasoning mode that applies extended inference chains to hard analytical problems before returning a response. The full GA is not yet live; prediction markets put the odds of a June 30 release at roughly 50%. But the technical profile is defined enough to evaluate now.
What 2 Million Tokens Actually Unlocks
To make the number concrete: 2 million tokens holds roughly 1,500 pages of dense text, the full transaction history of a mid-size portfolio, an entire multi-file codebase, or several years of SEC filings for a public company — in a single inference call. Production frontier models have topped out around 200,000–500,000 tokens until now. That ceiling forced builders to chunk documents, run multiple passes, and stitch results together — accepting that the model couldn't see patterns spanning chunk boundaries. At 2M tokens, those workarounds become optional for most financial document use cases.
For fintech builders, this is direct. KYC document review, AML case investigation, cross-period financial statement analysis, and regulatory compliance checks all involve corpora large enough that chunking introduces seam errors. A single-pass 2M-context call removes that failure mode.
Deep Think and the Reasoning Layer
Deep Think is Google's implementation of the extended reasoning modes that other frontier models have shipped — additional inference chains before the final output, trading latency for accuracy on hard tasks. Applied to a 2M-token input, you can give the model an entire regulatory filing corpus and ask it to identify all clauses with material compliance risk, reasoning through each before responding. Prior context limits made this a multi-step orchestration problem requiring a retrieval layer and prompt chaining. With Gemini 3.5 Pro, it is a single well-structured prompt.
Reading the Delay
The GA slip is worth noting without over-interpreting. Google announced a June target at I/O and is missing it. Serving 2M-token contexts at production API scale requires infrastructure that doesn't always ship on the schedule the announcement implied. The delay suggests Google is stress-testing capacity before opening broadly — not that the model itself isn't ready. Builders whose roadmaps are gated on Gemini 3.5 Pro availability should plan for a Q3 timeline and build against Claude 3.7 or GPT-5 while the enterprise preview matures.
What This Changes for Builders
The 2M context window is not equally useful for every task — short-form use cases don't benefit, and the cost-per-token scales above Flash. The shift is in use cases where context span was the binding constraint. Financial document processing, compliance automation, and codebase analysis all hit that ceiling regularly. Builders who shelved those ideas because the context math didn't close should revisit them now.
The broader pattern is worth holding: frontier context windows have grown roughly 100x in 18 months. Architectural assumptions you made when designing systems for 10,000-token documents have likely become unnecessary constraints. At 2M tokens, the document processing stack gets simpler — fewer retrieval steps, less chunking logic, less post-processing to assemble results. That simplification has real engineering and cost value once the model reaches broad GA.