Google said Gemini 3.5 Pro would launch on July 17. It didn't. The model has now missed three deadlines: June at I/O, June 30 as the target GA, and yesterday's leaked July 17 date. As of today, Google has not published a launch date, a model card, benchmark data, or pricing. The prediction markets are pricing August 7 at 73 percent for the next attempt.
This is not a scheduling slip. This is a broken timeline.
What I notice first: the talent context makes the delay understandable but not survivable. Noam Shazeer, the Transformer co-author Google spent a reported $2.7 billion to bring back from Character.AI in 2024, announced in June he was leaving again, this time for OpenAI. Nobel laureate John Jumper left for Anthropic the following day. Jonas Adler and Alexander Pritzel, both senior contributors to Google's AI coding and pretraining efforts, followed within the same week.
That sequence is not coincidental. The model was rebuilt from scratch after engineers found structural failures in recursive tool-calling and SVG generation. The re-pretraining cycle is what pushed the June date into July, and now July into August.
The real problem is not that Gemini missed a date. The real problem is that when it matters most, when you need stability and focus to ship a flagship model on time, the people who know the model best are leaving. Shazeer built this architecture. Jumper built components of it. When they go, you don't just lose engineers. You lose institutional memory of what the system is supposed to do.
I'm watching to see if Google even publishes independent benchmarks when Gemini 3.5 finally ships. If the model has truly been rebuilt twice, the internal eval numbers may not track anymore. Third-party benchmarks are how you rebuild credibility after a miss like this. Without them, every performance claim will carry the weight of three blown deadlines.
The talent exodus is the story Google won't say out loud but everyone in the field sees. You can't spend $2.7 billion to poach an Transformer co-author from a startup and then watch him leave for your closest competitor three years later without admitting something went wrong culturally. Shazeer had options. He chose to leave. Three days before Gemini 3.5 Pro and the Shanghai World AI Conference collide on July 17, the money, the silicon, and the geopolitics of AI are moving faster than the models themselves. Google is losing the race on all three fronts.
The leaked specs, if accurate, are still strong. Two million token context window. Deep Think reasoning on the paid tier. Pricing that undercuts OpenAI on output tokens. But specs don't ship themselves. People ship them. And right now, Google's people are walking out the door.
Top comments (0)