Andrej Karpathy Joins Anthropic: What the OpenAI Brain Drain Really Means
One of the most recognisable names in artificial intelligence has switched sides—and the move says something larger about where the centre of gravity in frontier AI research now sits.
On 19 May 2026, Andrej Karpathy quietly posted a short note on X: “Personal update: I’ve joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D.” No press release. No fanfare. Just a co-founder of OpenAI confirming he had taken a seat at OpenAI’s most credible rival.
For an industry that lives and dies by who is building which model, this is not a footnote. It is, depending on who you ask, either a routine mid-career move or the latest data point in a years-long pattern that should be deeply uncomfortable for OpenAI’s leadership.
Who Andrej Karpathy Is, and Why This Matters
Karpathy’s CV is short, dense, and difficult to dismiss. He completed his PhD at Stanford under Fei-Fei Li, then joined OpenAI as a founding member in late 2015—part of the original cohort alongside Sam Altman, Greg Brockman, Ilya Sutskever, John Schulman and a handful of others. In 2017 he was poached by Tesla to lead its Autopilot computer vision and Full Self-Driving programmes, a role he held until 2022. He returned to OpenAI in 2023 for roughly a year, then left in 2024 to start Eureka Labs, an AI-native education venture.
Along the way, he coined the phrase “vibe coding”—now shorthand for writing software in fluid collaboration with a large language model rather than line-by-line. He has educated a generation of practitioners through his YouTube course Neural Networks: Zero to Hero, which remains one of the most widely cited free resources for understanding how modern LLMs work from first principles.
At Anthropic, Karpathy will work on the pretraining team under Nick Joseph, focused specifically on using Claude to accelerate pretraining research. In plain language: AI helping to train the next generation of AI. That focus matters. Pretraining is the expensive, foundational phase that gives a model its underlying knowledge and reasoning. Putting one of the field’s most experienced practitioners on the problem of automating it is a clear bet on where Anthropic believes the frontier is heading.
This Is Not a Breakup. It’s a Pattern.
What gives the Karpathy move its weight is the company it keeps. Anthropic itself was founded in 2021 by a group of former OpenAI senior staff led by Dario and Daniela Amodei, who departed over directional differences—particularly around the commercial intensity that followed OpenAI’s $1 billion Microsoft investment. Anthropic’s DNA, in other words, has always been OpenAI exhaust.
That pipeline has not slowed. The last two years alone include three of OpenAI’s most senior departures, all ending up at Anthropic for at least some period:
- Jan Leike, OpenAI’s former head of alignment and co-lead of the Superalignment team, resigned in May 2024 with the now-famous parting line that “safety culture and processes have taken a backseat to shiny products”. Within two weeks he had announced his move to Anthropic to continue alignment research.
- John Schulman, an OpenAI co-founder and one of the architects of ChatGPT, left for Anthropic in August 2024. He stayed roughly five months before departing again in February 2025 to join Mira Murati’s Thinking Machines Lab—so Anthropic was a stepping stone rather than a final destination in his case, but the initial signal still pointed the same way.
- Andrej Karpathy, in May 2026.
The picture sharpens further when you zoom out from individual headlines to industry-wide hiring data. According to SignalFire’s 2025 State of Tech Talent Report, engineers at OpenAI are roughly eight times more likely to leave for Anthropic than the reverse. For Google DeepMind, that ratio is nearly 11:1 in Anthropic’s favour. Over the same period, Anthropic posted an 80% two-year retention rate, the highest among large AI labs; OpenAI’s sat at 67%.
Eight-to-one and eleven-to-one are not turnover ratios you can explain away with “Anthropic is the buzzy new startup”. They describe a one-directional flow.
What’s Actually Driving It
The convenient narrative is “safety culture”. And for some departures—Leike’s most overtly—that explanation clearly applies. But the fuller answer is more interesting, and it cuts across several dimensions.
Mission alignment as a hiring filter. Anthropic’s head of go-to-market recruiting, Nick Lewis, has been explicit that the company is willing to reject top-tier candidates if they don’t resonate with its mission. The result is a workforce that didn’t join chasing maximum compensation, which in turn makes them harder to poach back.
Compensation discipline. CEO Dario Amodei has publicly refused to engage in the multi-million-dollar signing-bonus wars that Meta has tried to wage. On the Big Technology Podcast in mid-2025, he framed it as a question of fairness: paying a randomly courted engineer ten times what an equally talented colleague earns would corrode the culture that drives retention in the first place. The counterintuitive result is that Anthropic loses fewer people than firms paying far more.
Research autonomy. Multiple former-OpenAI hires have cited the freedom to set their own research agenda at Anthropic as a meaningful pull. That is precisely the framing Karpathy used in his own announcement—”get back to R&D”—suggesting his decision was less about leaving Eureka Labs (he says he plans to resume that work in time) and more about returning to deep research at a lab where he can do it.
Belief in the underlying technology. Karpathy has been studying and teaching the mechanics of how LLMs are trained for the better part of a decade. When someone of his calibre signals that the next few years at the frontier “will be especially formative”, and then picks Anthropic to spend them at, that is not a neutral observation. It is a bet on which lab is most likely to move the field forward—and on which models will be the substrate that other models are trained against.
The Strategic Read for OpenAI
It would be a mistake to write OpenAI off. The company still has the largest paying user base of any AI lab, deep commercial integration through Microsoft, and a brand that remains synonymous with ChatGPT in most consumers’ minds. Its revenue keeps climbing. But the talent flow is, at this point, a strategic problem of its own kind.
The deeper concern for Sam Altman’s team isn’t any single departure—it’s that the people most likely to know what frontier AI looks like from the inside keep voting with their feet, and they keep voting the same way. Two of Karpathy’s fellow OpenAI co-founders—Schulman and now Karpathy himself—have spent time at Anthropic in the last 22 months. Ilya Sutskever, another co-founder, left to start his own safety lab, Safe Superintelligence Inc., in 2024.
That isn’t yet a death knell. But it is a signal—one which OpenAI’s leadership will need to answer in some way other than another product launch.
What This Means for the Wider Industry
For anyone outside the immediate orbit of the frontier labs—the engineers building voice, video and communications products on top of these models—Karpathy’s move has a quieter implication. Pretraining research is what determines, over the next few cycles, which lab’s models become the default substrate that everyone else builds on. The choice of where the best researchers spend their time will, indirectly, shape every downstream developer’s stack.
At SoftPage we have been tracking this question from several angles. In All the AI apps: same foundations, unique results? we looked at why models trained on broadly similar data still diverge in personality and capability—a question that becomes more pressing when AI is increasingly being used to train AI. How do free AI models make money? sketches the commercial pressures that drive the very behaviour Leike resigned over. And our AI archive covers Anthropic’s earlier Claude Opus 4 behavioural research, AI content moderation, and the ongoing question of which jobs survive the curve.
For anyone building communications infrastructure specifically—WebRTC, SIP, softphones, browser-based calling—the AI talent war matters because it sets the cadence at which natively-AI features (real-time transcription, intent recognition, voice-driven routing) reach production. The labs that win the long pretraining game decide the floor for everyone else.
The Closing Thought
Andrej Karpathy joining Anthropic is, in isolation, one researcher moving between two well-funded companies. Read alongside the SignalFire data, the Leike resignation, the Schulman detour and the broader founding story of Anthropic itself, it is something else: the latest entry in a steadily accumulating record showing where the people who have built frontier LLMs from scratch now want to spend their time.
That doesn’t predict who wins. But it tells you who the most informed observers in the field are betting on.
Sources: TechCrunch, CNBC, Axios, SignalFire 2025 State of Tech Talent Report, Wikipedia: John Schulman, Wikipedia: Jan Leike.
