At Google I/O in May, Sundar Pichai stood on stage and promised Gemini 3.5 Pro would arrive in June. A month later, it’s still not here. The new target: July 17. The reason? Google decided to run a fresh pretraining run instead of reusing the old Gemini 2.5 Pro foundation .
That‘s not the typical Google AI story we’ve been writing for the past two years. This time, they‘re not rushing. They’re not shipping a half-baked product to meet a self-imposed deadline. They‘re taking extra time to get it right.
The delay has also given Google time to integrate feedback from Gemini 3.5 Flash users — particularly around token consumption and reasoning performance. The additional weeks are being used to refine the model’s instruction-following and long-context capabilities .
The Context Window Arms Race
The technical specs alone would have made this release notable. According to leaked information from Reddit, Gemini 3.5 Pro supports a 2-million-token context window — enough to process an entire enterprise codebase in a single prompt .
For developers, this shift is significant. A 2-million-token model eliminates the need for RAG pipelines in many use cases — you can drop an entire repository into the context window without chunking, retrieval, or the architectural complexity those patterns demand. The trade-off is cost: a full 2-million-token input would consume roughly 4 times the tokens of a 500K context model, making it expensive for routine use but potentially cost-effective for large-scale analysis tasks .
Frontend Generation That Finally Competes
The benchmark data is impressive, but the real signal is in developer reactions. Leaked screenshots show Gemini 3.5 Pro generating polished SVG assets and frontend designs that one developer called “some of the best work I‘ve ever gotten from an AI model.” In head-to-head comparisons, testers reported that 3.5 Pro significantly outperforms previous Gemini models on frontend layout, design composition, and visual execution.
One simple prompt generated four distinct, detailed robot SVGs — a marked shift from the “lazy” Gemini of previous generations. The improvement appears to come from the new pretraining run and optimized architecture that Google is keeping under wraps.
Nano Banana Pro: A New Image Contender
The same 3.5 Pro foundation is also powering an updated version of Nano Banana Pro, Google’s text-accurate image generation model . Early expectations suggest the new model will compete directly with GPT-Image 1, with improved text rendering, higher resolution (2K/4K), and the ability to maintain consistency across up to five people. If these claims hold, Google may have a serious challenger in the image generation space.
Why the Delay Matters
Google‘s decision to run a new pretraining run instead of shipping on schedule suggests internal confidence that the model’s ceiling is higher than initially expected. This mirrors the pattern we saw with OpenAI‘s GPT-5.6 — the most impactful releases are often the ones that took longer than promised.
The timing is also strategic. GPT-5.6 is expected to launch on July 7, just as Claude Fable 5 users lose access to their free tier quotas. Google is betting that by July 17, the dust will have settled — and Gemini 3.5 Pro will be ready to capture the users who are shopping for a new model.
P.S. Google delayed Gemini 3.5 Pro because it chose quality over calendar. The 2-million-token context puts it in a league of its own. Frontend generation is finally competitive. July 17 is the date to watch. If it delivers on early testers‘ claims, Google has a genuine edge — and a much-needed win. But delivery has always been Google’s biggest challenge. This time, they‘re setting themselves up to meet it.
