Ex-Anthropic startup Mirendil raises $1 billion Series A

Mirendil, a startup founded by two former Anthropic researchers, has raised a $1 billion Series A led by Kleiner Perkins and Clearlake Capital. The company announced the round on Thursday, Nov. 12. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also invested. Mirendil did not disclose its valuation.

The round lands five months after a $200 million seed at a $1 billion valuation. Andreessen Horowitz and Kleiner Perkins co-led that seed, with Nvidia joining. In September, reports said Mirendil sought up to $1 billion at a $5 billion valuation with Kleiner Perkins leading. The announcement confirms the size and the lead. It does not confirm the price.

What does Mirendil claim to have built?

Mirendil's pitch is that AI agents can now do most of the hands-on work of an AI lab. A small team can build what once took an institution. "When agents do the hands-on work, the size of your team stops limiting what you can build," the company wrote.

Its central evidence is Mirendil-Base, a coding model it published results for on Oct. 21. Mirendil says a single researcher, working with agents, pretrained the model from scratch in 3.5 months. All-in compute cost came to less than $10 million. That figure, the company says, includes the agents' token spend, data processing and every small-scale experiment. Agents wrote more than 99% of the code, ran experiments and diagnosed failures. The researcher decided what to try and what to trust. Mirendil says it used only publicly available data, with no distillation and no licensed datasets. These are the company's own figures. They have not been independently replicated.

Mirendil then post-trained the base model into Fern. The company says Fern holds up to 10 million tokens in context. That equals roughly half a million lines of code, or days of an agent's work at once. Fern 0.1 is listed at 0.1 to 0.5 trillion parameters, with access "coming soon." Two larger models, Mirendil-Base-Large and Mirendil-Base-XL, sit at up to 2 trillion and 10 trillion parameters. Both remain in pretraining. Mirendil corrected two figures in its Fern post on Oct. 27 to "more accurately reflect" its training results.

What is Lor, the product it plans to sell?

Lor, now in private beta, is the platform Mirendil uses internally and intends to sell. It runs hundreds of agents at a time. The agents write code, run experiments and iterate for days. They provision their own GPUs and split jobs across dozens of sub-agents. Lor works with outside frontier models, including Anthropic's Claude Opus and Claude Fable, OpenAI's GPT-6 Sol and GPT-6 Astra, and open models such as Kimi, DeepSeek and GLM, as well as Mirendil's own. In one example the company cites, 102 agents built training environments from 100 PyTorch pull requests in an average of eight minutes each.

Who is behind the company?

Chief executive Neyshabur and Mehta met at Google in 2019. Both moved to Anthropic in late 2024 and left in December 2025. Neyshabur spent more than five years at Alphabet co-leading reasoning research for Gemini. Mehta built the first version of Anthropic's internal AI-research platform. Mirendil launched with roughly 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind and OpenAI. It still calls itself "a small team."

Why the stakes are high

The company frames its goal as handing frontier AI research to scientists and engineers outside the big labs. That places it in the most contested part of AI development. Anthropic has described recursive self-improvement, in which AI systems speed up their own development, as a potential danger. Mirendil's founders call it the shortest path to faster science and a risk that can be supervised. Thinking Machines Lab is chasing a related market; last week it released Inkling 2 Small, a model built to run AI training jobs.

Mirendil has now raised $1.2 billion in five months on unreplicated benchmarks. The next test is whether Lor's public launch converts agent-run research from a claim into a product with paying customers.