Andrej Karpathy, a co-founder of OpenAI and former head of Tesla’s Autopilot and Full Self-Driving programs, announced on May 19 via X that he has joined Anthropic. He will lead a new team focused on using Claude itself to accelerate pretraining research — the computationally intensive phase that gives large language models their core knowledge and capabilities. Karpathy will work under Anthropic’s pre-training lead Nick Joseph, with the company confirming the hire and stating that the team will leverage Claude to speed up its own development pipeline.
The move brings together one of the most recognized researchers in AI with a company that has been aggressively challenging OpenAI’s dominance. Anthropic also simultaneously hired cybersecurity veteran Chris Rohlf for its frontier red team, signaling a dual investment in capability and safety. Rohlf, previously with Meta and Yahoo’s Paranoids team, said AI can dramatically improve cybersecurity and that Anthropic is the ideal place to pursue that.
Karpathy’s career path has been closely followed. He left OpenAI in 2017 to lead Tesla’s Autopilot, returned briefly to OpenAI in 2022, then founded AI education startup Eureka Labs in 2024. He stated he remains passionate about education and plans to resume that work later. The hiring underscores a broader AI trend: competition is shifting from raw compute scaling to the quality and innovation of research teams.
Anthropic appears to be gathering momentum. The latest Ramp AI Index shows it overtook OpenAI in corporate adoption for the first time, reaching 34.4% of businesses in April versus OpenAI’s 32.3%. Over the past year, Anthropic quadrupled its business adoption while OpenAI’s grew just 0.3%. The company is reportedly raising fresh capital at a valuation near $950 billion, exceeding OpenAI’s $852 billion, on the strength of revenue growth, the Claude Code developer tool, and the cybersecurity-focused Mythos model.
Still, challenges remain. Anthropic has faced complaints about outages, rate limits, and a recent model update that triples token costs for image-containing prompts. Meanwhile, cheaper open-source inference services are growing fast, and OpenAI’s Codex remains a low-cost alternative. The hire does not end the race, but it shifts the burden of proof: the question is no longer whether anyone can catch OpenAI, but whether OpenAI can hold its lead.
