Two Ex-Anthropic Researchers Raised $200 Million With No Product. The Valuation Is $1 Billion.

Mirendil came out of stealth on June 25 with $200 million in seed funding and a $1 billion valuation, which makes it one of the largest seed rounds anyone has announced in AI. Andreessen Horowitz and Kleiner Perkins co-led it, with Nvidia also putting in money. The company was founded by Behnam Neyshabur and Harsh Mehta, who both left Anthropic in December 2025 after about a year there. It has not shipped a product.

That last part is the story. A billion-dollar valuation usually comes with revenue, or at least software people can log into. Mirendil has neither yet. What the investors bought is a team and a bet.

What they say they are building

Mirendil's pitch is AI that does the work of an AI researcher: designing experiments, running hyperparameter searches, evaluating models, and iterating on training runs with less human input each cycle. The plan, as the founders describe it, is to train models that specialize in those research tasks and then package them as a platform other organizations can point at their own problems.

The example the company uses is a university biology lab that wants a model for drug-target identification but has no machine-learning engineers. Instead of hiring a team, it would hand the problem to Mirendil's system and let it run the experimental loop. The promise is to compress research cycles that take months of human work into days of automated iteration.

That is a clean story. It is also, right now, only a story. There is no released model, no benchmark, and no methodology anyone outside the company can check. Mirendil says it will put out a first model and product in the coming months. Until then, the claim that its systems can do frontier research is exactly that, a claim.

What a billion dollars buys before launch

Strip out the product and you are paying for the people. Neyshabur spent about five and a half years at Alphabet, where he worked on Gemini's reasoning research, before joining Anthropic at the end of 2024. He and Mehta met at Google in 2019. The founding group runs to roughly 20 researchers and engineers pulled from Anthropic, OpenAI, Google DeepMind, and xAI, including Shayan Salehian, an early member of xAI, and Tara Rezaei, an MIT graduate.

Investors have been writing this kind of check for a while now.

Figures via Tech Funding News.

Most of these companies sold investors on a team and a direction rather than a working product. Safe Superintelligence has raised $6 billion at a $32 billion valuation with nothing commercial to show. Mirendil is the smallest of the group by valuation, and it is aiming at a narrower target: the research-automation layer that could, in theory, sit underneath all the others.

The part that is genuinely unsettled

The thesis runs straight into one of the oldest arguments in the field. If AI can meaningfully speed up AI research, the work compounds. Better models help build better models, and the advantage the large labs hold (thousands of researchers, years of accumulated infrastructure) starts to erode. Plenty of people find that prospect exciting. Plenty of safety researchers find it alarming, because a system that improves itself with less human input is also a system that is harder to supervise.

Mirendil's founders treat the safety questions as engineering problems, solvable with enough oversight, rather than reasons not to build. That is a defensible position. It is not a settled one, and stating it confidently does not make the hard part easy.

What is concrete here is the money and the resumes. Whether the underlying idea works is what the $200 million is there to find out. If it does, the round will look cheap. If research automation turns out to be five years away instead of one, $200 million is a very expensive way to fund a hypothesis.