When a company loses four senior researchers in six days, something systemic is happening. When three of those four land at a single competitor — and that competitor is Anthropic — it’s worth paying attention.
What happened#
Between June 18 and June 24, 2026, Bloomberg confirmed four departures from Google DeepMind that together represent one of the most significant talent redistributions in AI history:
- Noam Shazeer (June 18): Transformer co-author, led Gemini’s architecture through multiple generations, went to OpenAI
- John Jumper (June 20): Nobel laureate in Chemistry for AlphaFold, joined Anthropic
- Jonas Adler (June 24): Google DeepMind’s AI coding lead, joined Anthropic
- Alexander Pritzel (June 24): Pretraining specialist, AlphaFold contributor, joined Anthropic
Three out of four to Anthropic. Two of those three — Jumper and Adler — in areas directly tied to Anthropic’s core work: reasoning science and AI-assisted coding.
This follows Andrej Karpathy’s departure from independent research to Anthropic in May, where he joined the pre-training team. The pattern isn’t random.
Why they’re leaving#
The stated reason across multiple Bloomberg sources: GPU and compute resources are being reallocated inside Google from researcher projects to product deployments. DeepMind researchers are watching their experimental compute budgets shrink as the company prioritizes Gemini’s commercial rollout over open-ended research.
Anthropic and OpenAI are offering something Google currently can’t match: a place where frontier research still drives compute allocation rather than product KPIs — plus pre-IPO equity that makes the financial trade-off easy.
Shazeer’s destination (OpenAI) is notable because it suggests this isn’t purely an “everyone hates Google” story. Both Anthropic and OpenAI are absorbing Google’s talent. The split: OpenAI got the Gemini architecture lead; Anthropic got the Nobel laureate, the coding AI lead, and a pretraining specialist.
That split arguably favors Anthropic on the dimensions that matter most for Claude Code’s trajectory.
The Gemini 3.5 Pro connection#
At Google I/O on May 19, Sundar Pichai made an unusual public commitment: “Give us until next month.” He was referring to Gemini 3.5 Pro’s general availability.
June passed. As of June 26-29, multiple sources — CryptoBriefing, Investing.com, Startup Fortune — confirmed the model will not ship in June. Google cited the need for “more time with early testers to improve performance on complex, multi-step tasks.”
The timing is uncomfortable: the delay lands in the same week Google lost its Gemini co-lead and its AI coding lead. Whether those departures directly affected the 3.5 Pro timeline is impossible to know externally, but the optics are bad regardless.
Gemini 3.5 Pro’s specifications remain compelling on paper: 2M token context, Deep Think reasoning mode, and pricing estimated at ~$15/$60 per million input/output tokens. The 3.5 Flash (May 19) already led on MCP Atlas (83.6%) and offered 4× faster output than other frontier models — 3.5 Pro is supposed to close the reasoning gap.
But “supposed to ship in June” and “delayed to July” is exactly the kind of execution stumble that compounds: each day of delay is a day where Anthropic’s Fable 5 and OpenAI’s GPT-5.6 continue to accumulate evaluation data and developer mindshare.
What Anthropic gets from these hires#
John Jumper ran AlphaFold from a 4M-parameter proof-of-concept to a Nobel Prize-winning system that solved protein structure prediction. The relevant skill isn’t biology — it’s building systems that work at the edge of what’s scientifically possible and making them robust enough to be used by millions of researchers. That’s directly applicable to Anthropic’s hardest problems in multi-agent reliability and model capability research.
Jonas Adler was inside Google’s effort to make Gemini useful for coding. He has direct knowledge of what makes a coding-specialized model different from a general-purpose one — data pipelines, evaluation methodology, the evals that actually catch real bugs versus the ones that look good on benchmarks. Bringing him to Anthropic’s side while Claude Code holds its benchmark lead (Fable 5 at 87.6% SWE-bench Verified) is significant.
Alexander Pritzel worked on pretraining and AlphaFold — the mathematics of making large models generalize rather than memorize. Pretraining quality is the foundation everything else builds on.
Together they extend what Karpathy’s arrival already signaled: Anthropic is building a research organization that can sustain its capability lead rather than just riding it.
The broader competitive dynamic#
Google’s situation illustrates a structural tension that no hyperscaler has solved cleanly: the research culture and the product culture within AI labs are increasingly incompatible. Frontier research requires tolerance for unproductive compute, long time horizons, and experiments that fail. Product deployment requires the opposite.
Anthropic and OpenAI don’t yet have this problem — they’re still research-first organizations where the product (Claude Code, ChatGPT) is the downstream of the research, not the constraint on it. That changes as they grow, but for now the cultural gap is Anthropic’s most durable recruiting advantage.
The exit of four DeepMind researchers in six days is not a company failing. Google still employs thousands of excellent AI researchers and has more compute than most countries. But it is a visible symptom of the tension — and when the symptoms include your Gemini architecture lead, your Nobel laureate, your AI coding lead, and your pretraining specialist all in the same week, the signal is hard to ignore.
Watch Gemini 3.5 Pro’s July launch closely. If it ships with the benchmark numbers Google has implied, the talent story becomes a footnote. If it slips again, or ships with numbers that trail Fable 5 further than expected, the structural argument becomes much harder to dismiss.
Sources: Bloomberg talent confirmations June 24, 2026; CryptoBriefing; Investing.com; Startup Fortune; Bind AI