After missing a June go-live and spending most of July in a limited enterprise preview, Google has now confirmed a target launch date for Gemini 3.5 Pro: July 17. The reason for the extended wait turned out to be more than routine tweaks. Google scrapped the original base model entirely and rebuilt from scratch, targeting specific performance gaps that early testers flagged.

The timing is notable. Gemini 3.5 Pro arrives days before DeepSeek V4 graduates from preview to stable release on July 24, and while the wider rollout of GPT-5.6 is still gated behind government review processes. Enterprise teams evaluating their AI model stack now have three major frontier releases converging in a ten-day window.

Why Google Rebuilt From Scratch

The original Gemini 3.5 Pro architecture, an evolution of the 2.5 Pro model, reportedly could not close three performance gaps identified through Vertex AI enterprise testing: mathematical reasoning, complex SVG scene generation, and overall image quality.

These are not benchmark categories chosen at random. Mathematical reasoning is a proxy for the kind of structured, multi-step logical work that enterprise users need for financial modelling, code generation, and data analysis. SVG generation quality signals how well the model understands and generates structured, relational information — an indicator of performance on complex document and diagram tasks. Image quality matters for multimodal workflows and customer-facing applications.

Rebuilding rather than fine-tuning is an unusual call. It extends the timeline significantly but avoids carrying forward an architectural limitation. Google’s framing internally has reportedly been that it was not willing to ship a model that would fall short in direct comparisons with Fable 5 and GPT-5.6, particularly after the company’s AI team has seen notable researcher departures over the past six months.

What Gemini 3.5 Pro Is Expected to Deliver

The specifications that have been circulating through enterprise preview channels are consistent: a 2 million token context window, a Deep Think reasoning mode for extended multi-step problem solving, and autonomous workflow capabilities designed to chain complex coding and tool-use tasks.

The 2 million token context window would double the 1 million cap on Gemini 2.5 Pro. At that scale, the practical use cases shift. You can process an entire company contract library in a single call, ingest a year of customer support conversations alongside product documentation, or reason across a full codebase without chunking. These are not incremental improvements — they change the category of problem you can actually solve with a single model session.

On pricing, Gemini 3.5 Pro is expected to come in at $15 per million input tokens and $60 per million output tokens, making it a premium-tier model. Deep Think reasoning access will be gated behind the Ultra subscription tier at $250 per month.

Until independent evaluators run long-context retrieval benchmarks on the generally available model, these specifications remain third-party reporting and enterprise tester accounts rather than officially confirmed numbers. The figure to watch is not whether the model accepts a 2 million token prompt, but whether reasoning quality holds across the full range. Gemini 2.5 Flash users reported token efficiency issues in extended workflows — that is the bar the Pro rebuild needs to clear.

The Model Market Window: Ten Days, Three Players

The convergence of releases this month is unusual even by the pace of 2026.

July 17: Gemini 3.5 Pro general availability targeting July 24: DeepSeek V4 graduates from preview to stable release, and legacy API aliases retire Ongoing: GPT-5.6 Sol, Terra, and Luna remain in limited government-vetted preview, with broader access still pending

For enterprise teams that have been waiting to make a model commitment, this window is the one to evaluate in. The three models represent genuinely different approaches. Gemini 3.5 Pro is a dense transformer rebuild optimised for reasoning quality, visual precision, and long-context retrieval. DeepSeek V4-Pro uses a Mixture-of-Experts architecture — 1.6 trillion total parameters with 49 billion activated per token — giving it a strong cost advantage at $0.87 per million output tokens while targeting enterprise-scale throughput. GPT-5.6 Sol is built primarily for coding, biology, and cybersecurity tasks and is expected to reach broad availability in the coming weeks.

Cost profiles are diverging sharply. Gemini 3.5 Pro at $60 per million output tokens sits in the premium tier alongside Fable 5. DeepSeek V4-Pro at $0.87 per million output is 60 times cheaper. For teams making procurement decisions based on cost-to-complete rather than raw benchmark scores — which is how Microsoft and many enterprise buyers now evaluate models — the gap between those tiers demands a task-specific evaluation, not a default assumption.

What This Means for Business

If you are on Google Cloud or Vertex AI: Request Gemini 3.5 Pro preview access now if you have not already. Being in the early access group gives you a head start on integration and a chance to flag issues before GA. Enterprise teams with contracts in progress should ask their Google account manager about pricing and access tier details.

If you are evaluating models for long-context document work: The 2 million token context window is the single most important differentiator to verify. If it delivers as advertised, use cases around contract analysis, financial document processing, and large codebase reasoning become significantly more practical. Validate on your actual documents, not on benchmark datasets.

If you are cost-sensitive: Gemini 3.5 Pro at premium pricing is not a replacement for DeepSeek V4-Flash or V4-Pro for high-volume tasks. These models are playing different positions. Use the two-week window to test your specific workflows against V4 while it is in stable preview, before DeepSeek’s GA release on July 24.

If you are managing an AI model stack: Three major frontier model changes in ten days is exactly the scenario that makes informal AI management painful. Knowing which of your workflows is on which model — and having a lightweight process for validating output quality when a model changes — is no longer optional at scale.

The frontier model market is delivering genuinely useful tools. The challenge is building the internal competence to choose between them, configure them correctly, and catch quality regressions when provider infrastructure shifts.

That capability — reading the landscape, evaluating tools rigorously, building for maintainability — is exactly what Enterprise DNA’s data and AI courses develop. For organisations that want strategic support making sense of the current model landscape, Omni Advisory brings a fractional AI advisor who can help you build a model evaluation and selection framework that holds up as the market keeps moving.

July 17 is nine days away. Start your Vertex AI preview request this week.