Why Google Rebuilt Gemini 3.5 Pro From Scratch Mid-Cycle

While OpenAI and Anthropic sprinted out of a US government review in early July, Google went the other direction: Gemini 3.5 Pro slipped. Teased at I/O in May, the model missed a June I/O target and then a 30 June general-availability date, and as of 9-10 July 2026 it remained in limited Vertex AI enterprise preview. Reporting now offers a cause — and it is the most consequential detail of the whole episode. According to leaks surfaced around 10 July, DeepMind did not merely fine-tune a late build. It abandoned the original 2.5 Pro base and ran an entirely new pretraining pass, effectively rebuilding the frontier model from scratch mid-cycle.

Pretraining a frontier model is the single largest fixed cost in the business — the compute, the data pipeline, the months of wall-clock time. To scrap a base that far into a release cycle and start the run over is not a routine schedule slip; it is a deliberate decision to eat that cost twice. Labs do not absorb that lightly. The revealing part is what it implies about the alternative: DeepMind apparently judged that shipping the existing 2.5 Pro-derived build was worse than paying to retrain. That reads less like a stumble and more like a bet that the ceiling of the old base was too low to compete with what OpenAI and Anthropic were about to ship, and that only a clean pretraining run would move the frontier far enough to matter.

The leaked specifications, if accurate, show where Google intends to compete. The headline figure is a 2-million-token context window — reportedly the largest of any production frontier model, and a natural extension of the long-context lead Gemini has held. Reasoning is where the tiering gets pointed: Deep Think, the model's extended-reasoning mode, is said to be gated to the $250/month Ultra tier, positioning the strongest capability as a premium product rather than a default. Reported API pricing sits near $1.25 input / $10 output per 1M tokens.

The table below sketches the reported shape of the release. Treat every cell as leaked, not confirmed.

Read together, these choices describe a model built to win on depth rather than breadth: a very long memory for enterprise document and codebase work, and a deliberate reasoning mode reserved for buyers willing to pay for it. The Vertex-first, enterprise-preview rollout reinforces the read — Google is aiming this build at the accounts that value context length and are least sensitive to a premium reasoning tier.

The most quietly significant claim in the reporting is a competitive one. Gemini 3.5 Pro was framed as the only major frontier model set to release without US government restrictions — reportedly because its offensive-security benchmark scores sat below the unofficial threshold that triggered scrutiny for OpenAI's GPT-5.6 and Anthropic's Fable and Mythos models. If that inference holds, the second-order effect is striking: Google's slower, later, twice-retrained model gets to ship freely precisely because it is rated less dangerous on cyber-offense. The delay and the clean-release status may share a root cause — a model that is not pushing the offensive-security frontier is both less constrained by Washington and, plausibly, why DeepMind felt it had to retrain in the first place.

The strategic calculus, then, is subtler than "Google fell behind." A restriction-free release is a genuine go-to-market advantage — enterprise and international buyers wary of export controls or usage caveats get an unencumbered frontier model. But it is an advantage earned by scoring lower on the exact axis regulators watch, which is an awkward flag to plant. The rebuild suggests Google concluded the trade was worth it: pay to retrain, ship a cleaner, longer-context model a few weeks late, and turn the resulting compliance profile into a selling point. This story also threads off the DeepMind talent questions we covered on 25 June — a reminder that execution risk at the frontier is now as much organizational as it is technical. The tell will be the confirmed benchmarks: if a 17 July launch materializes with real numbers, we will finally see whether the second pretraining run bought Google the frontier position it paid twice for.