On stage at Google I/O on May 19, 2026, CEO Sundar Pichai looked out at a room full of developers waiting for one specific announcement and made them a promise: Gemini 3.5 Pro, the company's next flagship AI model, would arrive within a month. "Give us until next month to get it to you," he said. That month came and went. So did the next one.
What actually went wrong, according to Bloomberg
A Bloomberg report by Julia Love and Davey Alba, drawing on ten current and former Google employees, traced the holdup to a specific technical problem: Gemini 3.5 Pro's coding performance was falling short of rivals, particularly Anthropic and OpenAI's latest models. In late June, engineers updated the training data in an attempt to fix it — and the results, according to Bloomberg's sourcing, were disappointing. A separate report from HackerNoon described something more drastic: DeepMind scrapped the model's underlying base architecture entirely just days before a planned deployment, restarting on a heavier pre-training run built on a native Gemini 3 foundation rather than patching the existing one.
A Google spokesperson's public statement to Bloomberg was carefully narrow: the company is "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners," and is "shipping quickly across a wide range of models while keeping them highly cost-effective." Notably absent from that statement: a release date.
A market reading the delay as a bigger signal
The financial reaction has been sharper than the technical story alone would suggest. Alphabet shares fell 4.4% on the Thursday Bloomberg's report published, erasing roughly $200 billion in market value — less than four weeks after the stock had already lost an estimated $225 billion when senior DeepMind researchers departed for Anthropic and OpenAI in June. Neither sell-off tracked any change in Alphabet's underlying financial results: the company's most recent quarter showed $109.9 billion in total revenue, with Google Cloud growing 63% year over year. The market, in other words, is pricing in a narrative about Google's AI competitiveness, not a change in the numbers.
What Google shipped instead
Rather than nothing, Google used its July 21 announcement window to release three new Gemini Flash models and confirm — almost as an aside — that it has begun pretraining an entirely new flagship model, Gemini 4. Pretraining is the foundational stage of building a model, done before any fine-tuning, and confirming it has started signals Google is committing serious compute to its next real flagship rather than only fixing the current one. Gemini 3.5 Flash, already generally available, has performed credibly on its own — beating the older Gemini 3.1 Pro on coding benchmarks at roughly four times the speed, according to developer-focused coverage from Bind AI.
Why "Pro," specifically, is the one that matters
Gemini 3.5 Pro's specs, as confirmed so far, include a 2-million-token context window — double Claude Opus 4.8 and most other models generally available — plus a "Deep Think" reasoning mode aimed at long-document analysis and codebase-level reasoning. Those are the capabilities enterprise customers on Vertex AI are currently testing in early access, while the wider developer community waits on a release date Google has stopped publicly committing to. Whatever ships, when it ships, will be measured against a full year of Google publicly missing its own timeline for it — and against the researcher departures that, unlike the delay itself, may say more about Google's ability to hold onto talent than about any single model's benchmark scores.