Five months. That’s all it took for this AI lab to go from a $200 million seed round to a $1 billion Series A. Most startups need years to raise that kind of money. This one had about 20 people and a product still in private beta. So, what is Mirendil? Mirendil is a San Francisco AI lab started by two former Anthropic researchers, Behnam Neyshabur and Harsh Mehta.
It builds models and AI agents that do the work of AI research itself, like running tests and training models. The goal is simple: let any scientist build custom AI without a big team. In this guide, you’ll see who’s behind the company and what it has built.
You’ll also learn where the money came from and which claims still need proof.
Key Takeaways
- The company raised a $1B Series A on November 12, 2026, led by Kleiner Perkins and Clearlake Capital. That’s five times its June seed round.
- Founders Behnam Neyshabur and Harsh Mehta left Anthropic in December 2025.
- Its first model, Fern 0.1, claims a context window of up to 10 million tokens.
- Lor, its agent platform, runs hundreds of agents for days at a time. However, it’s still in private beta.
- Most performance claims come from the company itself, and outsiders haven’t tested them yet.
What Is Mirendil? The Short Answer
It’s an AI lab that wants AI to do AI research. In other words, it builds tools that design tests, train models, and check results, so people don’t have to do every step by hand.
The company calls its mission making AI R&D easier to reach. According to its Series A announcement from November 2026, it wants frontier research tools “within reach of every scientist and engineer.
That’s a bold promise. After all, building a top-tier model today takes giant teams and giant budgets. Here’s a real-world way to picture it. The Next Web described a university biology lab that wants a model to locate drug targets.
Normally, that lab would need its own machine learning team. With Mirendil, the idea is that AI agents handle that work instead. Neyshabur sums up the plan as “AI for AI for science.”
| Quick facts | Details |
| Legal name | Mirendil Inc. |
| Headquarters | San Francisco, California |
| Founded | 2026 (founders left Anthropic in December 2025) |
| Founders | Behnam Neyshabur (CEO), Harsh Mehta (CTO) |
| Team size | About 20 at launch |
| Total raised | $1.2 billion (seed plus Series A) |
| Products | Mirendil-Base, Fern 0.1, Lor (private beta) |
What most guides miss: This startup isn’t trying to beat OpenAI at chatbots. Instead, it’s selling the research process itself. That’s a different business, and it’s why investors see it as more than another model maker.
Who Founded Mirendil? The Anthropic and Google Roots
Behnam Neyshabur and Harsh Mehta started the company. Both are former Anthropic researchers, and both spent years at Google before that.
Their paths line up closely. The pair met at Google in 2019. Then they joined Anthropic in late 2024 and left in December 2025, as The Next Web reported in June 2026. Neyshabur, now CEO, spent more than five years at Alphabet. While there, he co-led reasoning research for Gemini.
Mehta’s background may matter even more. At Anthropic, he built the first version of the company’s internal AI research platform, sometimes working alone. In short, he’s already built a tool that helps a top lab run its research. Mirendil looks like an attempt to turn that idea into a product anyone can buy.
The early team is just as notable. It includes about 20 researchers and engineers from Anthropic, xAI, Google DeepMind, and OpenAI. For example, Shayan Salehian, an early xAI member, is part of the founding group.
Our take: Talent like this explains much of the valuation. Investors are betting on people who have seen how frontier labs work from the inside. Still, a great résumé isn’t a finished product. If you want to see how one of those labs builds its tools, read our guide to [Claude AI and how it works] (INTERNAL: Claude AI explainer).
Fern, Mirendil-Base, and Lor: What the Lab Has Built So Far
So far, the lab has shown three things: a base model, a long-context model built on it, and an agent platform. Only the model is open to early users, and most results come from the company itself.
Here’s how the three fit together.
| Product | What it is | Status (Nov 2026) |
| Mirendil-Base | A coding model pretrained from scratch | Announced October 2026 |
| Fern 0.1 | Mirendil-Base after post-training, with up to 10M tokens of context | Early access queue |
| Lor | A platform that runs hundreds of agents for days | Private beta |
The Mirendil-Base story is the most striking. The company says one researcher, working with agents, pretrained it in 3.5 months. Total cost stayed under $10 million.
According to MarkTechPost’s October 2026 report, AI agents wrote more than 99% of the code. If that holds up, it’s a strong proof of concept for the whole company.
Fern 0.1 launched on October 26, 2026. Its headline feature is a context window of up to 10 million tokens. The company says that’s about 23 MB of code, or roughly half a million lines. As a result, a model could read a huge codebase, training logs, or agent traces in one go.
Lor is the component that unifies everything. The startup says users can run hundreds of agents that write code and run tests for days at a time. It also says its team uses Lor every day.
The trade-off most coverage skips: A large context window doesn’t guarantee good reasoning across it. Startup Fortune reported that the company’s own multi-hop accuracy figure was 45% at full length.
That’s useful, but it isn’t magic. So treat the 10M number as storage, not proof of understanding. Furthermore, the $10M compute figure covers only pretraining, not the full cost of building the company.
Inside the $1B Series A: Who Invested and Why
The $1 billion Series A closed on November 12, 2026. Kleiner Perkins and Clearlake Capital led it, and the startup didn’t share a valuation.
Other named investors include Battery Ventures, Doerr Capital, Gil & Co., Google, M13, Pear VC, and S32, according to the official announcement. Several existing backers also joined. Meanwhile, the round came less than five months after the seed.
Here’s the full timeline so far.
| Date (2026) | Milestone |
| June 24 | $200M seed at a $1B valuation, co-led by Andreessen Horowitz and Kleiner Perkins, with Nvidia joining |
| August 6 | Multiyear Google Cloud deal worth more than $100M |
| October 26 | Fern 0.1 and Mirendil-Base announced |
| November 12 | $1B Series A led by Kleiner Perkins and Clearlake Capital |
The seed round alone was rare. SiliconANGLE reported in June 2026 that it valued the company at $1 billion. Later, in September, Tech in Asia reported that the lab was in talks to raise to $1 billion at a $5 billion valuation. However, the company hasn’t confirmed that figure. So treat $5 billion as a report, not a fact.
The detail worth watching: Google plays two roles here. First, it sells the startup compute. TechCrunch reported in August 2026 that the deal gives Mirendil access to Google’s TPUs, Nvidia GPUs, and managed training clusters.
Now Google is also an investor. That’s common in AI today, but it means part of the investment likely flows back to Google as cloud spending.
Furthermore, note what the announcement leaves out. It doesn’t say how the money will be spent. If you’re new to how these rounds work, our beginner’s guide to investing in startups explains seed, Series A, and valuation basics.
Can AI Really Speed Up AI Research?
The company’s core bet is that AI can improve itself faster than people can on their own. That idea is called recursive self-improvement, and it’s both promising and debated.
In simple terms, self-improving AI is a system that gets better at its job over time. Neyshabur explained the vision to TechCrunch with a simple image.
You can “point a problem at it, and it keeps getting better with time.” He gave Alzheimer’s research as one example. Other likely targets include medicine, biology, and materials science.
The upside is easy to see. Today, frontier research is slow and costly. A single test can take days. If agents can conduct hundreds of tests simultaneously, a small team could achieve significantly greater coverage.
The Mirendil-Base story, with one researcher in 3.5 months, is its best argument so far.
Yet the idea has critics. Anthropic has flagged recursive self-improvement as a possible danger if models change their code without oversight. The founders disagree.
They call it the shortest path to faster science, and they say it can be supervised. Gizmodo’s coverage captures the split well: some see a risky path, while others see a new way to do research.
Where we land: The safety debate is real, but it isn’t the main near-term risk for this startup. The bigger risk is proof. Currently, there are no public benchmarks for Mirendil-Base’s coding skills. Until outsiders are able to test Lor and Fern, their bold claims remain unverified.
What Mirendil Means for Founders, Researchers, and Investors
This lab matters most to people who need custom AI but can’t hire a full research team. For everyone else, it’s a signal of where AI tools are heading.
Here’s how different readers might approach it.
| Reader | Why it matters | Practical next step |
| Research labs | Could build custom models without an ML team | Join the Fern early access queue and test it on your data. |
| AI founders | Shows how small teams can now raise large rounds | Study how the founders sold the research process, not just a model. |
| Investors | Shows a new category: AI that does R&D | Watch for public benchmarks and Lor’s wider release. |
| Developers | Long-context models may change how you handle big codebases. | Compare Fern with other long-context tools you already use |
A common mistake to avoid: Don’t rebuild your roadmap around unproven tools. For example, a startup might plan a product that depends on Fern reading its whole codebase.
That plan breaks if accuracy drops at full length. Instead, run a small pilot first and measure results on your own work.
There’s a wider lesson too. This raise shows that investors still pay a premium for elite teams with a clear thesis. Its 20-person launch team raised more than most companies raise in a decade. For a broader view of tools in this space, see our roundup of the best AI coding tools.
Finally, keep your expectations balanced. The startup could become core infrastructure for science. Or it could stay a promising lab with unverified results. The next six months of public testing will show us which path we take.
Conclusion
Most AI startups did not move this fast in 2026. Its founders bring rare inside knowledge from Anthropic and Google. Its pitch is straightforward: AI agents that manage the demanding research tasks, enabling small teams to create what previously required a lab.
However, the narrative has two perspectives. On one hand, the funding is real, the Google deal is real, and Fern is in early access. On the other, key results are self-reported. The 10M-token window comes with a 45% accuracy figure at full length, and Lor isn’t public yet.
Our view? Mirendil is one of the most important companies to watch in AI research tools, but it hasn’t proven its biggest claims yet.
FAQs
1. Who founded Mirendil?
Behnam Neyshabur and Harsh Mehta founded the company after leaving Anthropic in December 2025. Neyshabur is CEO and co-leader of Gemini reasoning research at Google. Mehta is CTO and built the first version of Anthropic’s internal AI research platform.
2. How much has Mirendil raised so far?
The company has raised about $1.2 billion in total. That includes a $200 million seed round in June 2026 and a $1 billion Series A in November 2026, led by Kleiner Perkins and Clearlake Capital.
3. What is Mirendil Fern?
Fern is the lab’s first AI model. Fern 0.1 is built on Mirendil-Base and claims a context window of up to 10 million tokens. That’s enough for about half a million lines of code. Access is via an early access queue.
4. What is Mirendil’s valuation?
The startup was valued at $1 billion in its June 2026 seed round. The company did not disclose its valuation for the Series A round; however, Tech in Asia reported that earlier discussions valued the company at up to $5 billion, a figure that the company has not confirmed.
5. What is recursive self-improvement in AI?
Recursive self-improvement means an AI system improves its abilities over time, often by running experiments and training new versions of itself. Its founders see it as a faster path to science. Critics warn it needs strong human oversight.

Tabassum Shaik is an Author, Researcher, and SEO Specialist with over 8 years of experience creating informative content on business, startups, entrepreneurship, marketing, technology, and digital trends. She specializes in researching industry trends and transforming complex topics into practical, easy-to-understand insights. Her goal is to help readers stay informed, learn new ideas, and make better business decisions.