Andrej Karpathy Joins Anthropic: OpenAI Co-Founder Lands on Pre-Training Team

Andrej Karpathy — one of the most recognized researchers in AI and a founding member of OpenAI — announced Tuesday that he has joined Anthropic, where he will lead a new effort focused on using Claude to accelerate pre-training research. He started this week, reporting into pre-training team lead Nick Joseph.

The move is the largest publicly disclosed talent transfer between the two frontier labs since Anthropic was founded in 2021, and it lands on the same day Google's I/O keynote tried to reset the narrative around AI competition.

Why This Is a Big Deal

Karpathy is not just a senior researcher. He is one of the small group of people whose names are recognized by people who don't work in AI. He was OpenAI's first hire after Elon Musk and Sam Altman founded it, ran AI at Tesla through the Autopilot years, returned to OpenAI to work on GPT-4, and then left in 2024 to start Eureka Labs, an AI-native education company.

In the years since, he has become the most influential AI educator on the open internet. His YouTube series on building GPT-style models from scratch is the de facto onboarding curriculum for a generation of ML engineers, and his commentary on training dynamics regularly moves the conversation inside the major labs.

Why this matters: the public perception of "who is winning AI" is partly a perception of where the best researchers want to work. Karpathy joining Anthropic, even into a single team, is the kind of signal that compounds — it makes other senior hires easier and reframes Anthropic's pre-training group as the place to be.

What Karpathy Will Actually Do

An Anthropic spokesperson told TechCrunch that Karpathy will start a team focused on using Claude itself to accelerate pre-training research. Pre-training is the large, compute-heavy phase that gives a model its core knowledge before any fine-tuning or reinforcement learning is applied.

The framing matters. "Using Claude to accelerate pre-training" is a research direction Anthropic has hinted at for the last year — using current models to design better training data, evaluate intermediate checkpoints, and propose architectural changes. If it works at scale, it shortens the time and cost of every future Claude version.

In his own statement, Karpathy wrote: "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. I remain deeply passionate about education and plan to resume my work on it in time."

That last sentence reads as a soft commitment that Eureka Labs is on pause, not shut down. Karpathy is treating the Anthropic role as a multi-year focus rather than a permanent exit from teaching.

The Talent Wars Just Got Hotter

Anthropic also confirmed Tuesday that it has hired Chris Rohlf, a 20-year cybersecurity veteran, to its frontier red team — the group that stress-tests advanced AI models against severe misuse threats. Rohlf came from Meta, where he spent six years, and before that from Yahoo's "Paranoids" security team.

Two senior hires on the same day, both from rival hyperscalers, fit a pattern Anthropic has been running for most of 2026. The company crossed OpenAI in business AI adoption last month — 34.4% of US businesses now pay for Claude, versus 32.3% for ChatGPT, per the latest enterprise survey VentureBeat reported. Talent typically follows momentum, and the momentum has shifted.

OpenAI is not standing still. Sam Altman has continued to make large equity-heavy offers to senior researchers across the industry, and OpenAI's IPO push has given it a new currency for retention. But the optics of a founding OpenAI member crossing the street in the middle of an IPO road show are difficult to spin.

Expert Perspectives

The reaction on X has been close to unanimous: this is a signal hire. Several researchers pointed out that Karpathy's public writing has been increasingly aligned with Anthropic's positions on interpretability, alignment, and the value of smaller, cleaner training runs over brute-force scaling.

Others noted the timing. Karpathy joins as Anthropic ramps the next generation of Claude pre-training runs on its newly secured TPU and Colossus capacity. Having a researcher of his caliber on the team for the start of those runs — rather than coming in after the fact — is the optimal moment to influence the recipe.

A skeptical read: Karpathy is a famously independent operator. He left OpenAI twice and ran his own company for two years. Whether he integrates smoothly into a structured pre-training org is a real open question, and his comment about returning to education suggests the role has a built-in time horizon.

What's Next

Anthropic has not said publicly which Claude version Karpathy's team will most directly shape. The most plausible answer is the Claude 5 generation, expected later in 2026 on the new training infrastructure. If his pre-training acceleration work shortens that cycle even modestly, the impact shows up directly in Anthropic's release cadence.

For OpenAI, the immediate question is retention. Several senior researchers there have been publicly weighing offers from competitors, and a high-profile founding member departing mid-IPO process tends to accelerate similar decisions.

The Bottom Line

Karpathy joining Anthropic is the kind of move that does not change a single benchmark on its own, but reshapes the perception of where AI talent is concentrating. For Anthropic, it is validation. For OpenAI, it is a warning shot in the middle of its IPO push. For everyone else, it is one more data point that the center of gravity in frontier AI research is no longer a single company.