Reviews · JUNE 25, 2026
Google slips Gemini 3.5 Pro to July as five researchers exit in a week
Business Insider reports the frontier model's GA moved from June to July 2026 over long-horizon and token-efficiency issues surfaced by Antigravity and LMArena testers, the same week Noam Shazeer and John Jumper announced exits to OpenAI and Anthropic.
Google pushed Gemini 3.5 Pro's general-availability date from June 2026 to July 2026, Business Insider reported Tuesday, citing long-horizon agentic behavior and token-efficiency issues surfaced by internal testers running the model in Antigravity, Google's development environment, and external evaluators on LMArena. It's a one-month slip on a flagship that Sundar Pichai had teed up at Google I/O in May, where he unveiled Gemini 3.5 Flash and called the Pro launch "next month."
The slip itself is unremarkable. Frontier labs routinely buy themselves four to six weeks to reduce regressions on the evaluations that markets and Chatbot Arena watchers actually look at. What makes the timing legible to investors is what happened around it.
On June 18, Noam Shazeer, Gemini co-lead, Google VP of engineering, and co-author of the 2017 "Attention Is All You Need" paper that defined the transformer era, left Google for OpenAI. A day later, John Jumper, the Nobel laureate behind AlphaFold, announced he was moving to Anthropic. Two further DeepMind researchers are reportedly headed to Anthropic as well. Five departures inside seven days, by 24/7 Wall St.'s count, and the names aren't interchangeable. Shazeer is the person Google paid $2.7 billion in 2024 to license Character.AI's technology and bring back. Jumper is the face of DeepMind's most legible scientific achievement to non-specialists.
The market priced it cleanly. On June 22, the Wall Street Journal reported Alphabet shares fell 5%, erasing $225 billion in market value. GOOGL is down 5.09% on the trailing week to $345.29, with a further 1.14% drop Thursday morning to $341.34. Polymarket contracts on the frontier-model race repriced in tandem.
The deeper read is structural. Google's bench, on paper, is the deepest in the industry, anchored by Demis Hassabis at DeepMind. But token efficiency and long-horizon agent behavior aren't problems you fix with org-chart depth. They're tuned by the same small group of researchers who understand the model's training dynamics intimately, and that group is now smaller than it was a week ago. Tuning a frontier model in public, while the people who would do the tuning relocate to Mission Street and the Embarcadero, is a difficult posture to hold for long.
Google has been here before, structurally. The 2024 Character.AI deal was itself a $2.7 billion concession that the talent market had moved faster than the org could retain it. Doing the same maneuver twice in two years, with Microsoft- and Amazon-backed competitors writing the offers, suggests the doom loop isn't the model. It's the comp sheet.
Sources
- https://www.analyticsinsight.net/news/is-google-delaying-gemini-35-pro-launch-to-july-for-further-testing
- https://www.investing.com/news/stock-market-news/why-is-alphabet-stock-sliding-today-93CH-4760501
- https://247wallst.com/investing/2026/06/25/5-top-google-ai-brains-bolted-in-7-days-as-gemini-falls-behind-and-alphabet-stock-is-feeling-it/
- https://startupfortune.com/google-delays-gemini-35-pro-to-july-as-talent-exodus-deepens-the-pressure-on-its-ai-ambitions/
- https://www.digitaltoday.co.kr/en/view/74900/google-delays-gemini-3-5-pro-launch-to-july-to-incorporate-early-feedback-real-world-data