DeepMind Uses Claude Code | Steve Yegge vs Demis Hassabis 2026

A post from Steve Yegge, a veteran software engineer with stints at Google and Amazon and a former CEO role at the Grab superapp, went viral on Monday, April 20, 2026. Yegge alleged that Google DeepMind, the company's elite AI research division, has been quietly granted exceptions to use Anthropic's Claude Code as their primary autonomous coding tool, while the remaining 40,000-plus Google engineers are restricted to internal versions of Gemini. Yegge's framing was pointed: the people building the AI are not using the AI they are building. Within hours, Google DeepMind CEO Demis Hassabis had responded directly on X. The exchange has since become a flashpoint in the broader debate over whether Google 's AI tools are genuinely competitive at the frontier.

Yegge's Claim | DeepMind Researchers Use Claude Because Gemini Can't Keep Up

Yegge's argument rested on a specific technical distinction. He alleged that DeepMind researchers have been granted use of Claude Code because their internal tooling, including what he described as Gemini 3.1 Pro, lacks the capability to handle complex multi-file agentic coding workflows. Claude Opus 4.7 benchmarks at over 80 percent on SWE-bench, the most widely cited measure of a model's ability to resolve real GitHub issues autonomously. Yegge's contention is that this gap is not marginal for researchers doing deep systems work, and that granting DeepMind exceptions to use a competitor product is an implicit acknowledgment of it.

He also challenged the framing Google executives used in their pushback. When the figure of 40,000 engineers using AI tools weekly was cited as evidence of strong internal adoption, Yegge countered that weekly usage is not meaningful adoption. His standard for genuine AI integration is daily multi-hour sessions consuming millions of tokens, the kind of continuous coding-agent use that defines professional workflows at companies like Cursor or Replit. He claimed that threshold is being met inside DeepMind only, and not with Google's own model.

Hassabis Responds | "Completely False and Just Pure Clickbait"

Demis Hassabis replied on X with more directness than he typically deploys publicly. His post read: "Maybe tell your buddy to do some actual work and to stop spreading absolute nonsense. This post is completely false and just pure clickbait." Addy Osmani, Director of Google Cloud, also pushed back, pointing to the scale of Google's internal agentic coding platforms and the depth of the 40,000-engineer engagement figure.

Neither Hassabis nor Osmani offered specifics about which models DeepMind researchers use for their individual coding workflows, or whether any exceptions to internal tool mandates exist for particular research teams. The denial was categorical but not technical, which left the core question Yegge raised, whether DeepMind has any systematic or informal access to Claude Code, unverified in either direction.

Why the Claim Landed | The Strike Team and Sergey Brin's Code Reviews

The reason Yegge's post gained traction so quickly is that it arrived in the same news cycle as reporting from The Information confirming that Google has assembled an internal AI coding "Strike Team" specifically tasked with closing the gap on Claude Code . The team is working on a project called Antigravity, an internal coding environment designed to give Google engineers the same agentic, multi-file autonomous workflows that Claude Code provides externally. The fact that a Strike Team exists to build a Claude rival inside Google undercuts the message that everything is fine with the current stack.

Separately, leaked internal accounts describe significant pressure from co-founder Sergey Brin, who has reportedly been conducting code reviews and pushing engineering teams to accelerate AI integration. Brin's direct involvement in day-to-day technical work is unusual for someone in his position and is widely read as a signal that leadership believes the gap between Google's internal tools and frontier models from Anthropic and OpenAI is a genuine risk rather than a PR problem.

The Real Question | What the Strike Team Admission Means

Whether or not DeepMind researchers have informally used Claude Code, the more significant story is the one Hassabis cannot deny. Google has committed resources to building an internal coding agent capable of matching Claude's SWE-bench performance because its current internal tooling does not match it. The Antigravity project exists because the gap is real. The Strike Team exists because the gap matters.

Yegge's viral post may be imprecise on the internal exception details and Hassabis may be entirely accurate that no formal DeepMind policy involves Claude Code. But the structural argument, that Google's public-facing AI infrastructure is not yet what its own best researchers would choose if given a free choice, is consistent with the documentary evidence of where the company is directing resources. The race to close that gap through Antigravity is the story that will matter most when Google's next developer conference arrives in May.