Why OpenAI's Top Researchers Keep Joining Anthropic

On May 19, 2026, the man who helped invent OpenAI joined the company that was built to escape it. Andrej Karpathy, co-founder, the teacher behind half the deep-learning tutorials you have ever watched, gone to Anthropic. Within an hour the developer internet wrote the same headline: third senior defection in two years. Leike, Schulman, Karpathy.

There is one problem with that headline. The number is wrong. Not the direction, the count. The flow of senior OpenAI researchers into Anthropic is far bigger than three, it runs almost entirely one way, and the reason behind it is not the thing everyone assumes.

How many OpenAI researchers have actually joined Anthropic?

Around eleven, not three. If you count named, senior OpenAI researchers who moved to Anthropic since late 2023, the list runs to roughly a dozen, including two original co-founders and the former head of the Superalignment team, the group OpenAI built to keep future superhuman models under control. Over the same period, the number of senior Anthropic researchers who left for OpenAI is zero.

That asymmetry is the whole story, and it is what the “talent war” label gets wrong. A war has two sides taking losses. This is one team scoring while the other never gets the ball back. Before we trust that count, though, look at the names, because a few of them complicate the clean version of the story.

OpenAI → Anthropic 11 named, senior researchers since late 2023, including two co-founders and the head of Superalignment

vs

Anthropic → OpenAI 0 documented senior moves the other way in the same window. Not a misprint

A war usually has two sides shooting. This one has a scoreboard that reads eleven to nothing.

The roll call: eleven names, two co-founders

The list, dated and attributed, starts in May 2024 with Jan Leike, the Superalignment co-lead who quit OpenAI publicly and accused it of letting “safety culture and processes take a backseat to shiny products.” He was at Anthropic within days. Then came a co-founder, a second co-founder, a quiet stream of his old Superalignment colleagues, and most recently Karpathy.

What jumps out is the concentration. Seven of the eleven worked on safety or alignment. Two more, Karpathy and Durk Kingma, are pre-training people, the researchers who run the giant training jobs that give a model its raw knowledge. This is not a random scattering of leavers. It is a specific kind of researcher, leaving for a specific kind of place.

The viral “three defectors” number is wrong

The “Leike, Schulman, Karpathy” framing is catchy and slightly broken. John Schulman, the co-founder who invented the training trick that made ChatGPT usable, did join Anthropic in August 2024. He also left five months later, in early 2025, for Mira Murati’s startup Thinking Machines Lab. So the popular Reddit post celebrating his “defection” is narrating someone who is no longer there.

That detail matters, because it is the first crack in the tidy story. The flow into Anthropic is large, but it is not a frictionless one-way pipe. People pass through. The honest version is messier than a clean headline, so treat any single count with suspicion, mine included.

Are they leaving for the money?

No, and Anthropic goes out of its way to prove it. When Meta started waving $100 million signing packages at frontier researchers in 2025, Anthropic CEO Dario Amodei publicly refused to match them. His words: “We are not willing to compromise our compensation principles, our principles of fairness, to respond individually to these offers.” If a researcher were chasing the biggest check, the destination would be Meta, not Anthropic.

The retention numbers back him up. On the same podcast, Amodei said Anthropic kept 80% of the people it hired over the previous two years, the highest mark among the major labs. OpenAI sat at 67%, and Meta, the one writing the giant checks, came last at 64%. You do not buy loyalty like that with money you are refusing to spend. Something else is doing the work.

Anthropic 80%

two-year retention, the highest of the frontier labs, while refusing to match nine-figure counter-offers

Google DeepMind 78%

a close second, with the deepest pockets in the field behind it

OpenAI 67%

losing roughly one in three of the cohort the others kept

Meta 64%

lowest of the four, despite the $100M+ packages doing the poaching

So why do they keep choosing Anthropic?

Two reasons, and they are not the same reason wearing different clothes. The first is mission. After OpenAI dissolved its Superalignment team in May 2024, weeks after Leike and chief scientist Ilya Sutskever walked out, that safety crowd reassembled inside Anthropic. Leike runs the alignment science team there now, and several of his old colleagues report to him. They literally put the band back together.

The second reason is newer and has nothing to do with safety. Karpathy’s farewell did not mention it once. He said he wanted to “get back to R&D,” echoing Schulman’s earlier line about returning to “hands-on technical work.” The thread connecting them is a growing suspicion that OpenAI has drifted from research lab toward consumer-product company. In early 2026, OpenAI ran a Super Bowl ad for its chatbot. Anthropic ran one too, mocking the idea of running the ad. That gap, in one frame, is the culture these researchers are walking toward.

Reason one Rebuild the safety team

"Safety culture and processes have taken a backseat to shiny products."

— Jan Leike

OpenAI dissolved its Superalignment team in May 2024. Leike now runs Anthropic’s alignment science org, staffed with the same OpenAI alumni who co-wrote its "Weak-to-Strong" paper. They put the band back together.

Reason two Get back to R&D

"I am very excited to join the team here and get back to R&D."

— Andrej Karpathy

No safety critique, not one word. The thread connecting Karpathy and Schulman is the suspicion that OpenAI stopped being a research lab and became a consumer-product company that occasionally publishes.

What makes Anthropic structurally different?

Anthropic was founded in 2021 by an OpenAI walkout, over the same complaint these new arrivals have: mission drift. So when Amodei built it, he wired the governance specifically to stop that from recurring. The mechanism is the Long-Term Benefit Trust, a body that holds a special class of stock and elects a growing fraction of the board. And its members own zero equity in the company, so they answer to the mission rather than the share price.

As of April 2026, the Trust appoints a majority of Anthropic’s board. Translation: if investors ever tried to pull the kind of boardroom coup that briefly ousted Sam Altman at OpenAI in November 2023, the Trust would outvote them. For people who watched OpenAI’s mission catch fire from the inside, that arrangement is not a footnote. It is the reassurance that the same thing will not happen to them twice.

The board Controls the company Majority of seats elected by the Trust since April 2026

Investors

Amazon, Google, and others put in the billions. Money buys compute and equity, but their votes do not grow into board control.

Capital in, capped say

Long-Term Benefit Trust

Holds a special class of stock that elects a rising fraction of the board. Trustees own zero equity, so they answer to the mission, not the cap table.

Majority of the board

If investors ever try to pull a November-2023-OpenAI move, the Trust outvotes them. That is the whole point.

Is Anthropic actually winning the talent war?

Yes and no, and the “no” is the part the headlines skip. Anthropic is not really winning a war with OpenAI. It is winning a war with everyone else who tried to start an alternative. Senior OpenAI departures do not pool at one rival, they scatter across at least four. Sutskever left to found Safe Superintelligence, still secret, still without a product, yet valued around $32 billion. Murati left to start Thinking Machines Lab on a record seed round. Meta hired seven-plus OpenAI people in one summer.

So the map is not a duel between two labs. It is a churning mesh of senior researchers reshuffling between five of them, looking for the one least likely to disappoint them. Anthropic is winning the slice that wanted to join an existing lab with safety built in. That is a win, just a narrower one than “Anthropic is eating OpenAI’s brain” suggests.

Source OpenAI senior departures late 2023 – 2026

distributes to ↓

~11 Anthropic

Senior researchers who wanted an existing lab with safety baked in. Two co-founders, the Superalignment lead, Karpathy.

7+ Meta Superintelligence Labs

Bought talent outright with $100M+ packages. Roughly 40% of its first roster was ex-OpenAI.

Ilya Safe Superintelligence

Sutskever’s lab. Still no product, yet raised $2B at a reported $32B valuation in 2025.

Mira Thinking Machines Lab

Murati’s startup, a record $2B seed. Has since lost ~13 of its founding team, two co-founders back to OpenAI.

And the current even runs backward in places: two Thinking Machines co-founders quit and rejoined OpenAI in January 2026. Musical chairs, not a clean duel.

We have watched this exact movie before

If this pattern feels unprecedented, it is not. The cleanest parallel is not Xerox PARC, it is Google Brain in 2015. Back then, senior researchers, including a young Ilya Sutskever and Karpathy himself, walked out of the dominant lab to join a focused, mission-stated challenger that promised to put research first. That challenger was OpenAI. A decade later the roles have swapped, and the same kind of people are making the same move away from it.

History says one of two things happens next. OpenAI follows Google’s path: keeps shipping, keeps printing money, gradually cedes the research frontier to the upstart. Or it follows Xerox’s, where the parent that funded the breakthroughs never recaptures the lead. Both endings are still on the menu, and which one arrives will be decided in the model releases, not the resignation letters.

2015

Google Brain → OpenAI

Sutskever, Karpathy, and others leave the dominant lab for a small, mission-stated challenger promising to put research first.

2024–26

OpenAI → Anthropic

The same kind of people make the same move, except the challenger now wires its governance to prevent the mission drift the OpenAI founders later complained about.

What developers should actually watch

So when the next “OpenAI researcher joins Anthropic” headline lands, and there will be one, probably soon, remember the real shape of it. The flow exists, and it runs almost entirely in one direction. And it is fragmenting across four labs, not concentrating at one. The talent map is evidence, not a verdict.

Because none of this is settled by who poached whom. It is settled by whether the next Claude beats the next GPT, and whether Karpathy’s bet, using Claude to accelerate the training of the next Claude, actually produces a jump. Everything else, the farewell posts and the Super Bowl ads and the LinkedIn musical chairs, is very expensive sports gossip. Watch the models.