Why Google held back a huge new AI model at its big conference

• Google delayed its Gemini 3.5 Pro AI model, disappointing some developers.
• The smaller Gemini 3.5 Flash model now powers Google's Antigravity AI coding service.
• Feedback from 3.5 Flash will likely be used to enhance Gemini 3.5 Pro, via reinforcement learning.

Google usually saves its biggest product launches for its I/O conference every year. This time, CEO Sundar Pichai held back, and it says a lot about where the company stands in the AI coding race.

During the keynote address, Pichai told the crowd that Google's new flagship Gemini 3.5 Pro AI model wasn't ready yet, drawing audible groans.

I was there, and I've spent the rest of the event coming up with a theory for the apparent delay: Google is holding this new model back for a while to get it even better at AI coding tasks.

Anthropic's Claude Code took the world by storm last year, and OpenAI's Codex has gotten much better recently. These frontier labs are scooping up developer mindshare and generating big revenue by enabling coders to use AI tools with agents to automate and speed up coding tasks.

It's a revolution that's upending Silicon Valley, and Google was probably a bit behind. But not for long.

Instead of releasing 3.5 Pro, Pichai spoke at length and with passion about another new model, the Gemini 3.5 Flash. This is a smaller model that's faster and a lot cheaper, while being only slightly less powerful than the world's current top models.

Google has already made 3.5 Flash the main model powering its Antigravity AI coding service. Starting today, software developers will use this tool to churn out code.

This will generate a mountain of anonymous and highly valuable data. For instance, if an engineer starts a new coding project in Antigravity and suddenly halts the task, it suggests that something in the output from Flash 3.5 wasn't right.

Google can use this feedback data to improve the larger 3.5 Pro model, likely through reinforcement learning—a technique in which a new AI model is refined by rewarding good outputs and punishing bad outcomes.

Signals from running Antigravity on the smaller 3.5 Flash model will likely help with this process in important ways. That's because coding is particularly good at generating clear signals for AI model development. If the code is good, it likely works. If it's bad, it often breaks stuff.

This should give the larger 3.5 Pro model strong clues about which coding outputs worked and which didn't.

"I know you can't wait to get your hands on it," Pichai said onstage. "Give us until next month to get it to you."

So, when this big new model finally lands, I expect it to be a lot better at coding — the hottest application of generative AI right now.

Sign up for BI's Tech Memo newsletter here. Reach out to me via email at [email protected].