On July 17, 2026, Google DeepMind released Gemini 3.5 Pro, its most capable model to date — and one of its most delayed, after engineers scrapped the original version and ran a full pretraining rebuild from scratch.
The result is a model built around two differentiators: a 2-million-token context window — double what Gemini 3.5 Flash offers — and a new Deep Think reasoning layer that enables multi-step logical backtracking at inference time.
Why Google Had to Rebuild It
The delay wasn't PR management. According to reporting from Startup Fortune and other outlets, Google's engineers discovered that the original Gemini 3.5 Pro failed systematically at recursive tool-calling — situations where the model calls a tool, uses the result to decide whether to call another, and so on. SVG generation was another consistent weak point. Rather than patch these in post-training, DeepMind ran a new pretraining round from scratch, pushing the launch from its original window into mid-July.
What 2 Million Tokens Actually Means
Two million tokens is roughly 1.5 million words. In practical terms, that's enough to hold an entire medium-sized codebase — or a set of lengthy legal documents, a full research library on a narrow topic, or a year's worth of meeting transcripts — in a single prompt without chunking. Chunking degrades quality: every time you split a document and process it in pieces, you lose cross-document context. A genuine 2M window sidesteps that entirely.
For developers building retrieval-augmented generation (RAG) systems or document pipelines, this is a meaningful architectural shift. It doesn't eliminate RAG as a pattern — latency and cost still favor selective retrieval at scale — but it removes the hard upper bound that forced workarounds.
Deep Think: Reasoning at Inference Time
Deep Think is Gemini 3.5 Pro's reasoning mode, available on the $99.99/month Ultra tier. Like similar reasoning features in competing models, it allows the model to work through a problem in multiple internal passes — pausing, reconsidering, backing up — rather than committing to a single forward pass.
The gains tend to show up most clearly on math, formal logic, and multi-step science problems. Whether that translates to real-world improvement on ambiguous, under-specified tasks is what developers will be stress-testing over the next few weeks.
Where It Sits Competitively
The enterprise preview pricing lands at roughly $12–15 per million input tokens and $36–60 per million output tokens — positioning it in the same bracket as Anthropic's Opus tier, not the mid-market. OpenAI's GPT-5.6 Sol competes on high-end reasoning at $5/$30 input/output per million tokens, making it the cost-efficient alternative for similar use cases.
Google's bet is that 2M context plus Deep Think is a defensible combination that neither Anthropic nor OpenAI exactly replicates today. Whether the market agrees will show up in adoption numbers over the next quarter.
Gemini 3.5 Pro is available via the Google AI API and Google AI Studio starting July 17.