Two weeks ago, this site covered Google’s own target of 17 July 2026 for Gemini 3.5 Pro’s general availability. That date came and went. On 21 July, Google shipped three other Gemini models — but not 3.5 Pro. Bloomberg’s reporting, since corroborated by TechCrunch, 9to5Google, Search Engine Journal and Neowin, gives the reason: coding performance kept missing Google’s internal bar. Google reportedly reset and updated Gemini 3.5 Pro’s training data in late June specifically to fix this, and the results were still disappointing enough to hold the release back further. Google DeepMind’s product lead, Logan Kilpatrick, has said the model is currently being tested with partners and the company hopes it will “land soon” — language that stops well short of a date.
Part of the delay looks organisational rather than purely technical: separate teams inside Google Cloud, DeepMind, and the Android division are reportedly building overlapping AI coding tools in parallel, competing for the same limited compute rather than pulling in one direction. That’s a structural problem, not a bug that gets fixed in one more training run.
So what
We got the July date wrong two weeks ago because we reported Google’s own stated target as if it were a shipping date — and this is the same mistake worth flagging for anyone planning a product roadmap around a specific frontier model’s release. Coding ability has become the single hardest capability for a general-purpose model to nail reliably, which is exactly why it’s the bar holding back Google’s flagship. If your AI feature plan currently assumes “we’ll swap in the new frontier model when it lands,” build the fallback path now rather than when the date slips a second or third time — model releases in this market are directionally predictable but rarely date-predictable. That’s the kind of model-agnostic planning we build into AI product development — get in touch if a roadmap decision is currently waiting on a release date that hasn’t happened yet.