Mirendil, a San Francisco startup building AI agents that automate scientific and engineering research, has raised $1 billion in a Series A round led by Kleiner Perkins and Clearlake Capital. The round, announced November 12, 2026, closed roughly five months after the company's $200 million seed round and brings its disclosed total funding to $1.2 billion. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also participated, alongside NVIDIA, which had backed the seed, according to Securities.io.
Mirendil was founded in early 2026 by Behnam Neyshabur and Harsh Mehta, both of whom left Anthropic in December 2025 after joining the company in late 2024. Neyshabur previously spent five and a half years at Google leading reasoning research for Gemini. The roughly 20-person team also includes alumni of xAI, DeepMind and OpenAI.
Why Agent-Led Research Is the Pitch
Mirendil's bet is that the bottleneck in frontier AI is no longer compute alone but the scarcity of researchers who can design experiments, tune hyperparameters and diagnose failed training runs. Its software, Lor, runs hundreds of agents in parallel to do that work. The company says agents wrote more than 99% of the code for one project, ran the experiments and diagnosed the failures themselves, with minimal human oversight.
The company points to a working proof point: one researcher used Lor's agents to pretrain Mirendil-Base, a coding-specialized model, from scratch in 3.5 months on less than $10 million of compute. That model was then post-trained into Fern, which the company says can hold up to 10 million tokens of context, enough to track days of an agent's work without losing the thread. Fern access is listed as coming soon, while Lor remains in an invitation-only private beta that is free during the beta period.
What the Money Is For
Mirendil has not disclosed a post-money valuation for the Series A, but the jump from a $200 million seed five months ago signals investors are pricing in rapid traction rather than a finished product. The company says proceeds will go toward extending the same model: letting a single person take on projects that once required a full team, and letting a team attempt work that once required an entire institution.
That requires more compute, more specialized models beyond Mirendil-Base and Fern, and enough engineering headcount to keep Lor's agent fleet reliable as it scales from single researchers to enterprise deployments.
Investors and Prior Rounds
Mirendil's seed round, roughly five months before this Series A, raised $200 million at a $1 billion valuation and was co-led by Andreessen Horowitz and Kleiner Perkins, with Nvidia also participating. Kleiner Perkins stayed on to co-lead the Series A alongside new entrant Clearlake Capital, a private equity firm not typically associated with early-stage AI research bets. The participation of Google and Nvidia as strategic investors mirrors a pattern across frontier AI labs, where cloud and chip suppliers buy into startups that will become large compute customers.
Competitors and How This Raise Compares
Mirendil sits in a crowded but young category of "AI for AI research" tools, alongside efforts inside OpenAI, DeepMind and Anthropic's own internal tooling, plus smaller independents building automated ML engineering agents. What sets Mirendil apart on paper is speed: a $1 billion valuation at the seed stage and a billion-dollar Series A within five months is a pace closer to the frontier model labs themselves than to typical developer-tooling startups. For the robotics and physical-AI sector, the relevance is indirect but real — if agent-led research shortens the cycle for training and fine-tuning large models, it could also shorten the cycle for the world models and vision-language-action systems that robotics companies depend on.
What to Watch Next
The test for Mirendil is whether Fern and Lor can move from private beta and a single flagship case study to repeatable results across outside research teams. Investors are betting on the trajectory of two researchers who spent roughly a year at Anthropic; whether that experience generalizes to a platform other labs will trust with their core research workflows remains unproven.