Mirendil Raises $1 Billion Series A Five Months After Its Seed, Betting One Researcher Plus Agents Can Do Frontier Work

Mirendil, a San Francisco startup founded by former Anthropic researchers to build tooling for frontier AI research, has raised a $1 billion Series A led by Kleiner Perkins and Clearlake Capital. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also participated. The company did not disclose its valuation, but the Wall Street Journal reported it at about $5 billion, roughly five times the value assigned in its seed round.

The pace is the story. Mirendil announced a $200 million seed on June 24, led by Andreessen Horowitz and Kleiner Perkins, with Nvidia joining later as an investor and a Google Cloud partnership following in August. That means roughly $1.2 billion raised in five months by a company that has not shipped a generally available product. Dealroom places the round among the largest early checks in AI, alongside Safe Superintelligence, Thinking Machines Lab and Recursive Superintelligence.

The team explains much of the willingness to write it. Chief executive Behnam Neyshabur co-led Google DeepMind's Blueshift team before moving to Anthropic, where he worked on long-horizon reasoning and research systems. Chief technology officer Harsh Mehta built the first version of Anthropic's internal AI research platform. Co-founders Shayan Salehian, previously at xAI, and Tara Rezaei, who interned at OpenAI, round out a founding group of about 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind and OpenAI.

The technical claim behind the fundraise landed in October. Mirendil said a single researcher, working with agents, pretrained Mirendil-Base, a coding model in the 100-billion- to 500-billion-parameter range, from scratch in about three and a half months for less than $10 million of compute, including preliminary experiments, the final run, agent tokens and data processing. Agents wrote more than 99 percent of the code, ran experiments and diagnosed failures; the researcher chose directions and judged results. The company says the work used only public code, pages and datasets, with no distillation or synthetic data. Mirendil has not published benchmark scores, so competitive coding model remains its own description.

Two products follow from that. Fern, post-trained from Mirendil-Base, carries a context window of up to 10 million tokens, enough to review days of an agent's working record, and is listed as coming soon. Lor, introduced on October 28, is a cloud platform where hundreds of agents write code, run experiments and provision compute on machines the user selects, working for days or weeks at a stretch. It is in an invitation-only, free private beta on macOS, Linux and the web, with Windows, iOS and Android promised later.

Mirendil frames all of this as access. The argument is that frontier research capability is concentrated in a handful of labs because clusters and headcount are expensive, and that agents can compress both, letting a university group or a small biotech run work that used to require an institution. The company lists drug discovery, chemistry, biology and robotics as target areas. It has not disclosed how the $1 billion will be spent beyond models, agents and infrastructure.

There is a sharper version of the same thesis, and investors are buying that too. A system that can meaningfully accelerate AI research is a system that can improve itself, which is precisely the capability some researchers warn could outrun human oversight. Mirendil sells it as the fastest route to AI that solves hard scientific problems. The timing is pointed: Anthropic, where much of the founding team worked, prices its IPO on November 18, days after this round closed.

What it means: the seed-to-$1-billion jump is a bet on a research result rather than a revenue line, and the test is reproducibility. If outside teams can repeat the one-researcher, sub-$10-million pretraining result on Lor, the cost floor for frontier-adjacent work drops and a second tier of labs becomes viable. If they cannot, Mirendil is an expensive bet on pedigree, and the industry will have learned that agent-driven research demos travel further than they scale.