Google Is Losing Its AI Bench to a Company That Isn't Even Public Yet. It Can't Out-Bid Pre-IPO Equity, and Anthropic Knows It.
Four senior Google AI people left in six days — most recently Gemini's coding lead Jonas Adler and pretraining researcher Alexander Pritzel, both to Anthropic. The driver isn't culture; it's arithmetic. Pre-IPO equity at a ~$965B lab headed for a $1T listing is a payday Alphabet stock, already priced for success, structurally can't match. The two threads the coverage misses: Anthropic is winning this drain (three of four), not OpenAI, and the damage is aimed at coding — the most contested battleground in AI — exactly as Gemini 3.5 Pro slips. The arbitrage also has an expiration date: once the labs go public, the lottery ticket becomes a blue chip too.
Google has a problem it cannot solve with money, which is a strange sentence to write about Google. Over six days, four senior AI people walked out — most recently Jonas Adler, who led Google's coding work on Gemini, and Alexander Pritzel, who worked on pretraining, both to Anthropic. The reason isn't culture, mission, or a fancier title. It's arithmetic. Anthropic is a private company valued around $965 billion and pointed at a trillion-dollar IPO, and pre-IPO equity at a lab priced to roughly re-rate in public markets is a payday that Alphabet stock — already priced for success — structurally cannot match. You don't out-bid a lottery ticket with a blue chip.
That's the part worth sitting with, because it makes this something other than the usual talent-drain story of a hot company poaching from a slow one. This is a structural arbitrage that exists only in a specific window — the months before two labs go public — and it happens to be pointed at the one capability Google can least afford to bleed right now. Quick correction on the framing first: the breathless version has this at five departures, and the cleaner accounting is four in six days. The exact tally is fuzzy and not the point. The pattern is the point, and the pattern is verified.
The window opened with a shock. When Noam Shazeer — one of the authors of the transformer paper that started all of this — left for OpenAI and Nobel laureate John Jumper left for Anthropic, Alphabet shed about $269 billion of market value in a single day. That number told every Google researcher with a pre-IPO offer exactly what their departure was worth as a headline, and it told Anthropic and OpenAI that Alphabet was, briefly, vulnerable. The Adler-and-Pritzel pair is the second duo to go to Anthropic in a week. Whether that's a coordinated recruiting operation timed to the vulnerability or just opportunism feeding on itself, I can't confirm — the targeting theory is suggestive and unproven, and I'd flag anyone asserting it as fact.
What's not fuzzy is the damage, because it's aimed. Adler didn't lead a generic research group; he led coding, which in 2026 is the single most contested battleground across every frontier lab — SWE-bench, Terminal-Bench, agentic coding, the capabilities that actually convert into enterprise revenue. He was losing that race for Google: Gemini 3.5 Pro is already delayed and stuck in limited preview. So Google didn't just lose a senior researcher; it lost the person it was counting on to close the coding gap, to the exact competitor that gap favors. Adler's expertise now compounds Claude's lead in the lane he was hired to defend. That's the difference between an unlucky exit and a strategically expensive one.
Here's the asymmetry almost nobody is naming: Anthropic is winning this. Three of the four recent departures went to the Claude maker, not to OpenAI. The lazy frame is "Google loses talent to its rivals," plural, as if the outflow were evenly distributed. It isn't. In the IPO-era talent map, one lab is converting Alphabet's stumble into a specific, coding-shaped advantage, and the other is mostly a footnote to this particular run. If you're tracking who comes out of the pre-IPO scramble with the strongest bench, the direction of these specific moves matters more than the count.
Now the discipline, because the bear case on Google can outrun the facts. Two researchers, however senior, are a signal, not a capability collapse — Google DeepMind is deep, and one coding lead leaving does not by itself reset the frontier. The genuinely unanswered question is what Alphabet has actually done since the $269 billion day: accelerated vesting, retention grants, a comp overhaul for frontier researchers? No concrete retention response has surfaced, and that silence is the variable that decides whether this stops or compounds. If Google has a counter and it's working, this is a bad week. If it doesn't, it's the start of something structural.
The thing to understand about this arbitrage is that it has an expiration date built in. The entire advantage rests on pre-IPO equity at a private valuation that's expected to jump when the company goes public. Once Anthropic and OpenAI actually list — Anthropic first, OpenAI weeks behind — that equity re-prices to public-market reality, the lottery ticket becomes a blue chip, and the magnet weakens. Which makes the real question a timing one: can Google's AI bench survive the next few months until the window closes, and does Gemini 3.5 Pro's coding quality visibly slip against Claude and GPT in the meantime? That product gap, if it shows up in the benchmarks, is the moment a talent-flow story becomes a competitiveness story. I think Google holds the institution and loses the sprint — it has the depth to survive the drain but not the instruments to stop it before the IPOs close the window. What would change my mind is concrete: a visible Alphabet retention package that stems the exits, or the flow reversing once the labs are public and the equity stops being magic. Until one of those happens, Google is in the uncomfortable position of being the richest company that can't pay enough.