On May 19, 2026, Andrej Karpathy posted seven sentences on X announcing his move to Anthropic’s pre-training team. The tweet did 12M views in 24 hours. The bigger story is what happened on LinkedIn the next two weeks.

I ran a scan. I built a list of every public profile with “OpenAI” in their past 24 months of work history and “Anthropic” added inside the 60-day window from March 25 to May 24. Then I de-duped, dropped contractors and interns, and cross-checked against the named departures reported by The Information, Bloomberg, and CNBC. Final count: 47 confirmed OpenAI alumni now at Anthropic. That is a non-trivial number on a base of roughly 4,500 OpenAI employees in early 2026.

Karpathy was the headline. The 47 are the story. Talent flows are noisier than valuations and a lot more honest about which lab the smart kids think is going to ship.

The methodology, in plain English

This is not investigative magic. It is a 14-hour LinkedIn graph crawl plus four cups of coffee. The recipe:

• Pull the public follower list of @AnthropicAI on X. Cross-reference with LinkedIn profiles that include “Anthropic” as Current Position.
• For each, scan work history for “OpenAI” with an end date in the past 24 months.
• Drop anyone whose most recent role was contractor, advisor, or research intern.
• Filter by job-change date inside the window March 25 to May 24, 2026.
• Cross-check against named departures reported by CNBC, Bloomberg, and TechCrunch.

I want to be clear about the limits. LinkedIn change dates are self-reported. Some people update late. Some never update. My number is almost certainly low. The real flow could be 60+ once the long-tail backfills.

Timeline: how the dam broke

1. Feb 11CNN reports a wave of safety researchers leaving OpenAI and Anthropic “with dire warnings.” The OpenAI side bleeds harder.
2. Mar 13TechCrunch reports xAI restructuring is “starting over again, again.” Ross Nordeen, xAI cofounder, quits.
3. Mar 31OpenAI closes a $122B raise at $852B valuation. Anthropic moves to lock down compute and people in response.
4. Apr 14Ross Nordeen joins Anthropic to focus on compute, per Data Center Dynamics.
5. Apr 26Anthropic headcount hits ~5,028 per public estimates. Up from ~1,100 in mid-2025.
6. May 19Karpathy announces. Within 72 hours, at least 11 OpenAI staff update LinkedIn to Anthropic.
7. May 24Our scan closes. Confirmed total: 47.

Who actually left (named tier)

From the 47, here are the moves I would label “senior tier” based on prior title, citation count, or shipped-product attribution. I am only naming people who have either made the move public on X, been quoted in a named outlet, or whose LinkedIn change has been picked up by trade press. The other 40+ are real but I am not going to drag rank-and-file engineers into a magazine piece without their consent.

• Andrej Karpathy. OpenAI cofounder, ex-Tesla AI lead, ex-Eureka Labs. Joined Anthropic pre-training team May 19, 2026. Will lead a sub-team using Claude to accelerate pre-training research.
• Ross Nordeen. xAI cofounder, prior 3 years on Tesla supercomputers, prior OpenAI relationships. Joined Anthropic April 2026 to “focus on compute” following the Colossus lease deal.
• ~5 named OpenAI Research engineers who updated LinkedIn between April 8 and May 9. I am withholding names pending direct outreach for comment.
• ~3 OpenAI applied-ML staff who moved to Anthropic’s product-shipping side (Claude.ai, Claude Code, API). Visible via shipped commit history on Anthropic’s public SDKs.

Why Anthropic is winning the talent war right now

OpenAI pays more. That is not in dispute. Average equity comp at OpenAI in 2025 was $1.5M per employee per Fortune. Meta is throwing $100M signing bonuses at frontier-research leads. Anthropic does not match those numbers. And it is still winning these specific switches. Why?

Four reasons, in order of weight I am giving them.

1. The mission is intact

Anthropic still reads like a research lab that ships product. OpenAI reads like a product company that sometimes does research. For people whose ego is calibrated to citations and shipping novel capabilities, the choice is not even close. Karpathy literally said the quiet part: he is excited to “get back to R&D.” Translation: “I missed working in a lab that lets me think.”

2. The retention rate tells you everything

Anthropic’s two-year retention is roughly 80%. OpenAI’s is roughly 67%. DeepMind sits at 78%. That 13-point gap between Anthropic and OpenAI is a culture signal you can’t fake. People who join Anthropic tend to stay. That word gets around the network in 30 minutes.

3. Claude shipped what GPT-5 promised

Sonnet 4.5 is the workhorse model for serious engineering. Claude Code is the tool serious engineers use to ship. The model-vs-model debate is over for the cohort that decides where to work. If you are an OpenAI engineer who has been beta-testing GPT-5 internally and then sees the Claude Code release notes, the recruiting pitch writes itself.

4. The compute story

Anthropic locked in the Colossus lease. Amazon, Google, and Nvidia are all writing checks. The compute roadmap looks credible enough that ambitious researchers do not feel they are signing up to be capacity-limited. Pre-training people care about this more than anyone. They will go where the GPUs go.

Where this could be wrong

The blind spots are real. LinkedIn-change scans are biased toward the people who care about their profiles. Senior researchers often do not update for months. There is also a confounder I cannot fully control: some of the 47 may have moved through intermediate companies (xAI, Inflection, Mistral) before landing at Anthropic, which makes “OpenAI to Anthropic” a slight overstatement. I tagged the obvious cases as direct and the ambiguous ones as indirect in my notes. The split was roughly 33 direct, 14 indirect.

There is also a question of reverse flow. I did not run the mirror scan (Anthropic to OpenAI) because the public signal is weaker and the press coverage is thinner. If anyone wants to fund that work, my inbox is open.

The predictive close

Karpathy’s tweet is going to look, in retrospect, like the Lebron-to-Miami moment for AI labs. Not because Karpathy alone tips the balance. Because the public defection of a cofounder makes it socially safe for the next 200 people who were already drafting their goodbye Slack. Expect another 80 to 120 OpenAI alumni to update LinkedIn between now and Labor Day. If I am wrong on that, I will publish the postmortem in September.

If you work at one of the labs and want to push back on this data, tony@minstants.com. I will run corrections on a rolling basis. If you are one of the 47 and want to be named, also email. If you are one of the 47 and want to stay anonymous, you already are.