Startup Founded by Ex-Anthropic Staff to Build AI That Improves Itself Is Already Improving Its Valuation
Mirendil has no public product, a mission statement about AI researching AI, and a valuation that went from $1 billion to a requested $5 billion in three months — which is either the market pricing in the future or pricing in the founders' old badges.
Three months ago, Mirendil was worth $1 billion, on the strength of a $200 million seed round and four names — two of them recent Anthropic alumni — with excellent LinkedIn pedigree and a plan to build AI systems that improve themselves with progressively less human involvement. This week, the company is reportedly in talks to raise up to $1 billion more at a $5 billion valuation, led by Kleiner Perkins, with Andreessen Horowitz also circling. That’s a fivefold jump in the time it takes most startups to finish onboarding their first ten employees, and Mirendil still has, per multiple outlets, no public product.
The pitch, as reported, is recursive self-improvement: models that plan their own experiments, run them, evaluate the results, and iterate — doing the job of an AI researcher, aimed at AI research itself. It’s the single most self-referential thing you can build in this industry short of an AI whose only task is fundraising, and Mirendil is not being subtle about the ambition. Co-founder Behnam Neyshabur co-led Google DeepMind’s Blueshift team before Anthropic; Harsh Mehta built Anthropic’s internal research platform; the founding four also include an early xAI engineer and a 23-year-old MIT graduate and Olympiad medalist, in case the résumé wasn’t already doing enough work. The company has more than 20 employees, a partnership with Google for Cloud TPU access, and, according to TechFundingNews, “no public product” — a phrase doing a great deal of quiet lifting in an article about a $5 billion valuation.
To be fair to the market, “no product yet” is not automatically a red flag in a sector where the entire premise is that frontier capability arrives suddenly and the winners are decided by who has the compute and the talent lined up when it does. Mirendil’s plan is to launch a frontier model supporting engineering and research work by early 2027, and it’s pitching itself as a way for university labs and materials-science teams to rent frontier-grade research capability without building their own AI infrastructure — a genuinely useful service, assuming it exists on schedule and assuming “AI that plans its own experiments” clears the bar of “reliable enough that a university physics department trusts it unsupervised.”
But the valuation math deserves a second look regardless. Five billion dollars is roughly 25 times the total capital Mirendil has raised to date, a multiple PYMNTS notes is “similar” to other frontier labs’ — which is less a defense of the number than an admission that the entire category prices this way now. TechFundingNews put the tension most plainly, asking whether the valuation “reflects genuine breakthroughs or merely the founders’ prestigious pedigrees.” Given that the company has, again, no public product, “genuine breakthroughs” is not currently on the table as an explanation. That leaves pedigree, timing, and a category of investor that has decided the safest bet in AI is founders who already worked at the company everyone agrees is doing it right.
There is a specific irony sitting on top of all this that’s hard to ignore: Mirendil exists because two of its founders left Anthropic, a company whose CEO spent September publishing essays arguing the industry should deliberately slow down the race toward more capable, less supervised AI systems. Mirendil’s entire business model is building AI systems that get more capable with less supervision, faster. Neither fact makes the other one wrong, exactly — “pace the frontier” was always a statement about what happens once a lab has decided to compete at all, not an argument against founding a new one. But it’s a tidy demonstration of how little a safety essay changes about where the talent and capital actually flow. Dario Amodei writes “we must pace the frontier” on a Saturday; by the following Monday, two of his former researchers are raising a billion dollars to build the thing the essay was worried about, and calling it democratization.
None of this requires Mirendil to be lying about its ambitions or its investors to be foolish about the math. Self-improving AI research systems might genuinely be the next unlock, and being early with the right team has been correct before. But a $5 billion valuation for a company with no shipped product, three months after its last valuation, built on the premise that AI can now be trusted to run its own R&D loop, is exactly the kind of number this industry produces when everyone involved has decided that skepticism is a luxury nobody can afford to be the one still holding when the round closes.
Sources: