Mirendil, a Frontier Lab Where Agents Do the AI Research, Raises $1B Series A

For years, building frontier AI has been the work of a few well-funded labs, with hundreds of researchers and enormous compute budgets behind each. Mirendil, founded by former Anthropic researchers, argues that premise is about to break. If AI agents do the hands-on research, it says, the size of your team stops limiting what you can build. On Nov. 12, the company announced a $1 billion Series A.

Mirendil is among the best-known of the “neolabs,” a new breed of frontier lab started by veterans of OpenAI, Google DeepMind, and Anthropic. They are structured more like research institutes than startups, and they are betting on breakthroughs that could define the next decade of AI rather than on near-term revenue. Mirendil’s bet is recursive self-improvement: handing the work AI researchers do today, such as designing experiments, fixing engineering problems, and improving AI systems, to AI itself. Co-founder Behnam Neyshabur has described the vision as “AI for AI, and AI for science.” The aim is to let scientists and engineers in fields like medicine and materials build specialized models of their own.

Handing a lab’s workload to agents

The product behind that idea is Lor, now in invite-only private beta. Think of it as a way to delegate to dozens or hundreds of AI agents the way you would to a team of researchers. A user sets a goal, and the agents write code, run experiments, and iterate on the results in the cloud, often for days at a stretch. When something breaks, they try to fix it rather than wait for a person. They request GPU machines on their own and release them when the work is done. Jobs keep running when a laptop is closed, and users can watch each agent’s progress and stop any of them.

Mirendil’s own examples show the range: running a batch of experiments overnight and writing up the results, auditing a repository for security bugs and opening a pull request for each fix, or reproducing a table from a research paper on eight GPUs and explaining where the numbers differ. In one test, 102 agents built training environments from 100 PyTorch pull requests, averaging eight minutes each. Lor works with models from Anthropic, OpenAI, Google, xAI, Meta, and others, alongside Mirendil’s own, and it is free during the beta. The company says it runs its own research on Lor every day.

Mirendil’s own model, Fern, supports up to 10 million tokens of context, roughly 500,000 lines of code. That is enough to read days of an agent’s logs in one pass to trace a failure, or to scan a training dataset for duplicates and low-quality data. Fern is built on Mirendil-Base, a pretrained model that one researcher, working with agents, trained from scratch in 3.5 months for less than $10 million in compute, according to the company. Agents wrote more than 99% of the code. Two larger models are still in training, and Fern is listed as coming soon, with an early-access form open.

A team from Google, Anthropic and xAI, with five times the capital in five months

Neyshabur led the Gemini team at Google DeepMind, where he drove pretraining for math and code reasoning, before leading a team at Anthropic. Co-founder Harsh Mehta also comes from Anthropic, and Shayan Salehian was an early member of xAI who worked on Grok. Mirendil still calls itself a small team.

Kleiner Perkins and Clearlake Capital led the round. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, and S32 joined, along with other new and existing investors. It comes five months after Mirendil’s $200 million seed round, and the valuation was not disclosed. Bloomberg reported in September that the company was in talks to raise up to $1 billion at a valuation of around $5 billion.

The competition: neolabs, and agent products

Mirendil joins a crowded field of labs founded by frontier-lab alumni. Reflection AI, started by former Google DeepMind researchers, raised $2 billion at an $8 billion valuation in October 2025 and released the open-weight model Beam last month, attacking the access problem by opening up the models themselves. Humans&, founded by alumni of Google and xAI, closed a $480 million seed round in January on a pitch of AI that collaborates with people rather than replacing them. Advanced Machine Intelligence raised $1.03 billion in a seed round in March, and London’s Ineffable Intelligence raised $1.1 billion in April to build AI that learns from its own experience rather than human data. Periodic Labs raised $300 million to apply AI to scientific experiments and automate materials design.

Where most of these labs lead with research and plan to ship later, Mirendil has put a beta product and a model in front of users first. In agent products, Cognition, maker of the AI software engineer Devin, raised more than $2 billion at a $48 billion valuation in September, and SpaceX acquired Cursor for $60 billion in an all-stock deal. Both focus on general software development, while Mirendil is aiming at AI research and scientific experiments. It also lets users run Anthropic’s and OpenAI’s models inside Lor, positioning itself as the layer that runs agents at scale across many models rather than competing head-on at the model level. Its results so far rest on the examples and figures the company itself has published.

Mirendil said the new capital will move it toward a world where one person can take on what used to need a team, and a team what used to need an institution.