Google Delays Gemini 3.5 Pro as DeepMind Loses Four Top Researchers

Google was expected to unveil Gemini 3.5 Pro back in June, in line with Sundar Pichai's announcement at the I/O conference in May. Instead of a launch, the company has announced yet another delay and, according to industry reports, is planning a full rebuild of the model from the ground up, abandoning the current architecture based on Gemini 2.5 Pro.

The reasons behind the delay, according to outlets tracking the AI industry, come down to three areas: token efficiency, code quality, and long tasks requiring multi-step reasoning. Internal testing reportedly showed the model falling short of the level Google publicly promised at I/O, particularly against upcoming models from OpenAI and Anthropic.

DeepMind's Wave of Departures

The departures piled on top of the delay at the worst possible moment. Noam Shazeer, one of the key figures behind Gemini's development, returned to OpenAI less than two years after Google paid $2.7 billion to bring him back along with the Character.AI team. Days later, John Jumper, who won the Nobel Prize in Chemistry for AlphaFold, left the company for Anthropic.

The week closed out with the departures of Jonas Adler, who led the team responsible for Gemini's coding capabilities, and Alexander Pritzel, a specialist in model pretraining. Both also moved to Anthropic. Four exits in six days from one of the world's leading AI labs rattled investors, especially coming right after the previously announced delay of the flagship model.

Market Reaction

Alphabet shares fell about 5 percent on the day the market learned the scale of the departures, wiping out roughly $225 billion in market capitalization. Looking at the broader trend, the stock slid from its May peak of $402.38 to $345.29 by late June, a drop of about 10 percent in six weeks. For investors, it's a signal that competing for top AI researchers is proving just as costly as competing on the models themselves.

New Architecture Instead of Patches

Rather than shipping a patched version of Gemini 2.5 Pro, the DeepMind team has opted to rebuild the model from scratch. The goal is to improve mathematical ability, SVG graphics generation, and overall image quality, areas where rivals have held an edge. It's a risky call, trading a longer delay for the hope of a more durable improvement in its standing against OpenAI's GPT-5.6 and Anthropic's Fable 5.

Some industry sources caution that the July 17 date is circulating in reports but has not been officially confirmed by Google. The company is keeping the model in limited testing access for select Vertex AI enterprise customers, promising wider availability in July without giving an exact date.

I'm excited to share that I'll be joining OpenAI - Noam Shazeer, former Gemini co-lead at Google DeepMind

What This Means for the AI Market

For companies relying on Google's models, the episode shows that a lead in the AI race can be fragile even for a giant with Alphabet's resources. Businesses planning Gemini-based integrations should expect schedule slippage, while rivals from OpenAI to Anthropic gain extra weeks to cement their latest models' positions before Google responds with an updated Gemini.

Researchers moving to direct rivals OpenAI and Anthropic further complicates matters, since the people leaving take with them knowledge of architecture and roadmap direction that competitors can now put to use in their own labs.

Sources: The Agent Report (the-agent-report.com), BigGo Finance (finance.biggo.com), Startup Fortune (startupfortune.com)