Google walks into I/O this week with a problem it can’t spin: its coding tools have fallen behind, and the company knows it. According to MIT Tech Review, the situation has gotten bad enough that Google reportedly had to let some DeepMind engineers use Anthropic’s Claude Code for their own work, just to keep them from falling further behind. That’s a remarkable admission from the company that gave us the Transformer.
This matters because coding ability has become the proxy metric for foundation model quality. When developers benchmark models, they benchmark code. When enterprises pick a stack, they pick based on what their engineers want to use. Right now, that’s Claude Code or OpenAI’s Codex, not Gemini.
The coding comeback play
MIT Tech Review reports that DeepMind has spun up a new AI coding team, and John Jumper, the Nobel laureate behind AlphaFold, is reportedly contributing to the effort. Expect a major coding announcement at I/O, likely an update to Google’s Antigravity agentic coding platform.
But temper expectations. Even Google’s internal models, which sit ahead of public releases, apparently weren’t good enough to stop engineers from fighting over Claude Code access last month. Two days of keynotes won’t close a gap that took 12 months to open.
Where Google still leads: science
Here’s the underreported angle. Google is the only frontier lab with a Nobel Prize on the wall. Its AI co-scientist tool has been called an “oracle” by a Stanford researcher quoted in MIT Tech Review, and AlphaEvolve keeps discovering new solutions to long-standing math and computational problems.
Science AI gets a fraction of the press that coding does, but the long-term commercial implications are enormous: drug discovery, materials science, protein engineering. If Google announces new scientific tools at I/O, they’ll matter more in five years than whatever coding demo lands the loudest applause this week.
Health is the wildcard. Google is publishing some of the strongest LLM-based health research out there, yet OpenAI has owned the consumer health conversation since ChatGPT Health shipped in January. Google’s new Health Coach goes public tomorrow, but it’s positioned around fitness and diet rather than medical advice. That’s either smart caution or a missed opportunity. We’ll know within a year.
The drama Google wants to avoid
While I/O runs in Mountain View, the Musk v. Altman trial wraps up 30 miles north in Oakland. MIT Tech Review notes that DeepMind’s Demis Hassabis has largely stayed out of the CEO mudslinging that’s defined the past quarter, presenting himself as a Nobel-winning nerd rather than a culture warrior.
Google isn’t drama free though. Last month 600 employees, many at DeepMind, signed a letter protesting an impending Department of Defense deal. Google signed it the next day. Expect Pichai and Hassabis to dodge those questions on stage. Expect the questions to surface anyway.
What practitioners should do now
Three practical takeaways from this moment:
- If you’re building with code-gen models, don’t switch your stack based on I/O announcements alone. Wait two weeks for independent benchmarks. Marketing demos and production reliability are different sports.
- If you’re working in science or health AI, watch Google’s announcements closely. This is where they’re genuinely ahead, and any new APIs or partnerships could reshape your roadmap.
- If you’re betting on a single foundation model, stop. The coding leaderboard has flipped twice in 18 months. Multi-model architectures aren’t a hedge anymore, they’re table stakes.
The broader story is that the foundation model race no longer has a single leader. Anthropic owns coding. OpenAI owns consumer mindshare. Google owns science. Meta owns open weights. Whoever tells you one company is winning isn’t paying attention.
Full coverage and Google’s I/O announcements are available at the original MIT Tech Review piece.