Gemini 3.5 Pro Lands July 17 With a 2-Million-Token Context Window

Google DeepMind will release Gemini 3.5 Pro to general availability on July 17, according to a leaked launch plan that circulated on July 10 and has since been corroborated by multiple outlets. The headline number: a 2-million-token context window — double anything currently shipping at the frontier, including OpenAI’s week-old GPT-5.6 family.

This is not an incremental bump. Gemini 3.5 Pro was reportedly rebuilt on an entirely new pretraining run rather than adapted from 2.5 Pro, the same approach Google took when it split the 3.5 Flash line out at I/O in May. A new “Deep Think” extended reasoning layer handles logic and math workloads — but Google is gating it behind the $250-per-month Ultra subscription, the clearest signal yet that frontier reasoning is becoming a premium SKU across the industry.

API pricing is the aggressive part. Leaked figures put Gemini 3.5 Pro at roughly $1.25 per million input tokens and $10 per million output tokens. That undercuts GPT-5.6 Sol by about 4x on input cost while doubling the context window. For teams doing long-document analysis, large-codebase work, or agentic pipelines that carry heavy context between steps, that combination is hard to ignore.

The timing is deliberate. July 17 is the same day DeepSeek plans to graduate its V4 family from preview to stable release, setting up a direct head-to-head between Google’s most capable model and China’s most-watched lab. And it comes barely a week after OpenAI’s GPT-5.6 launch and xAI’s public Grok 4.5 release — the most compressed frontier release cycle the industry has seen.

Google has not officially confirmed the date, but the leak’s specifics — pricing, tier gating, the Deep Think branding — line up with how the company has staged every Gemini release since 3.0. A launch-week surprise on benchmarks is the only real unknown left.

For developers, the practical takeaway: if you’re mid-migration to GPT-5.6 Terra or Luna for cost reasons, it’s worth waiting four days. A 2M-token window at $1.25 input changes the math for RAG architectures — some retrieval pipelines simply become “put the whole corpus in context” problems. That won’t be the right call for every workload, but for the first time it’s a real option at a defensible price.

Four frontier releases in ten days. The model-layer price war just found another gear.

Sources

Related reading