Thinking Machines co-founders make a dramatic return to OpenAI

The revolving door between OpenAI and its spin-off startups just took a wild, dramatic turn—complete with leaked memos, firing drama, and corporate espionage allegations.

On Wednesday, OpenAI announced that Barret Zoph and Luke Metz, both co-founders of Mira Murati's high-profile startup Thinking Machines Lab, are returning to the ChatGPT maker. They are joined by early staffer Sam Schoenholz in a move that blindsided the AI industry.

Fired on Wednesday, rehired by sunset

The split was anything but amicable. Tech reporter Kylie Robison first reported that Zoph had been fired from Thinking Machines for "unethical conduct." A source close to the startup alleged that Zoph had shared confidential company information with competitors.

OpenAI, however, painted a very different picture. In an internal memo to staff, OpenAI CEO of applications Fidji Simo explained that Zoph told Murati on Monday he was considering leaving, only to be fired two days later. Simo explicitly pushed back on the allegations, telling staff that OpenAI does not share Murati’s concerns regarding Zoph’s conduct.

Because the situation escalated so quickly, Simo noted that the hiring timeline was accelerated, leaving some role specifics to be sorted out later. What is known is that Zoph will report directly to Simo, with Metz and Schoenholz working under him. Murati quickly confirmed Zoph's exit on X, naming Soumith Chintala as the new CTO while completely omitting mention of the other departures.

A brutal blow to Murati's $12B venture

For Thinking Machines, this is an existential headache. The fledgling lab pulled off a massive $2B seed round at a $12B valuation last year, and was recently in talks to raise over $4B at a staggering $50B valuation. But capital cannot easily replace brainpower. Just this past November, another co-founder, Andrew Tulloch, defected to Meta.

For OpenAI, bringing Zoph and Metz back is a major coup, especially after recently losing VP of research Jerry Tworek. Before leaving in late 2024, Zoph served as OpenAI’s VP of post-training, overseeing the critical final touchups that turned raw models into ChatGPT and API endpoints. Metz spent two years contributing heavily to ChatGPT and the landmark o1 reasoning model. These are the exact engineers OpenAI needs right now.

Why it matters

In our view, this messy tug-of-war shows that the battle for elite AI talent has abandoned all pretense of Silicon Valley politeness.

For developers following Thinking Machines' main product Tinker—a tool aimed at helping developers customize AI models on proprietary data—losing the core post-training architects casts a shadow over the product's immediate technical roadmap. When the founders building the customization layer walk out the door, delivery timelines inevitably slip.

More broadly, this exposes the fragility of the "OpenAI mafia" startups. Massive valuations cannot shield young labs when key builders decide the frontier compute and deployment scale of a giant still outweigh founding equity. And OpenAI's willingness to brush off serious leak allegations to onboard Zoph at lightning speed sends an unmistakable message: when frontier capabilities are on the line, winning the talent war trumps everything else.