Google delayed the broader release of Gemini 3.5 Pro after internal testing found the model fell short of its own goals, particularly on code generation, per Bloomberg. The Pro model was announced alongside Gemini 3.5 Flash at I/O in mid-May and had been expected in June. Google updated the training data in late June to lift coding skills, and the results were disappointing, per 9to5Google. It remains in limited testing with corporate partners and the US government.
The detail worth sitting with is that a data-mix change failed to move coding. When a frontier model underperforms, labs have several levers: more compute, better data, longer post-training, different reinforcement objectives. Google reached for the data lever, retrained, and got a disappointing result. That points to something structural in the recipe rather than a shortage of tokens, and structural fixes take a full training cycle to test, which is why this delay runs in months rather than weeks. Coding is also the least forgiving place to miss, because it is verifiable. Either the code runs or it does not.
The timing compounds the problem. Anthropic shipped Sonnet 5 with near-flagship agentic coding at mid-tier pricing, and Moonshot released Kimi K3 as an open-weight model that developers preferred for front-end coding in blind testing. Google is holding its flagship back precisely as coding becomes the category models get judged on and where price pressure is fiercest. Engineers and managers inside the company are reported to be worried about losing ground to Anthropic and OpenAI. For buyers, the practical read is that Gemini 3.5 Flash is what you can actually deploy today.
Bottom Line
Missing on coding is the expensive place to miss right now. Watch whether Google ships Pro on the next training cycle or quietly repositions Flash as the workhorse.