🔥 Two Ex-Anthropic Researchers Just Raised a $1 Billion Series A, Five Months After Their Seed
Mirendil, a nine-month-old startup from two ex-Anthropic researchers, raised a $1B Series A on 12 Nov 2026, five months after its $200M seed.

A nine-month-old company with no generally available product just raised more money in one round than most startups raise in their entire lives. And it did it on the strength of two people, a few models, and a pitch that AI can now do the job of an AI lab.
Let’s get the number on the table first: $1 billion. That’s the Series A that Mirendil, a San Francisco startup founded in early 2026, announced on 12 November 2026, co-led by Kleiner Perkins and Clearlake Capital (Mirendil). The round also drew Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32.
Here is the part that makes investors’ eyes water. Mirendil’s $200 million seed closed back in June 2026 at a $1 billion valuation (The Next Web). So in under five months, the company went from a billion-dollar seed to a billion-dollar Series A, raising a cumulative $1.2 billion before shipping a public product.
The thesis is simple and slightly unsettling: if you can build an AI that does AI research, you can compound your way to everything else.
🧠 Why This Matters
Mirendil was started by Behnam Neyshabur and Harsh Mehta, and their CVs are most of the story. Neyshabur spent more than five years at Alphabet co-leading reasoning research for Gemini before joining Anthropic in late 2024. Mehta built the first version of Anthropic’s internal AI-research platform. The two met at Google in 2019, moved to Anthropic together, and left together in December 2025 (TechFundingNews).
That matters because the whole bet rides on their judgment about where model-building is going. Mirendil isn’t trying to out-chat ChatGPT. It’s building models that design experiments, tune hyperparameters, evaluate other models, and run training rounds with progressively less human input. The product is meant to be the researcher, not the chatbot.
“We’re building tools to democratize frontier AI R&D.” — Mirendil, Series A announcement
If that works even partway, it changes who gets to do cutting-edge AI. Right now, frontier research is gated behind a few labs with enormous teams and enormous compute budgets. Mirendil’s pitch is that a small team pointing agents at a problem can match them.
📊 Deep Dive
Mirendil spent the last few weeks putting out proof points instead of promises. Three of them anchor the round:
- Mirendil-Base — a competitive coding model that one researcher, working alongside agents, pretrained from scratch in 3.5 months for under $10 million in compute (Mirendil). Pretraining a capable base model is normally a job for a large team and a far larger budget.
- Fern — post-trained from Mirendil-Base, it holds up to 10 million tokens of context, which the company says is “enough for days of an agent’s work.”
- Lor — now in private beta, it runs hundreds of agents that write code and run experiments for days at a time. Mirendil says it runs its own work on Lor every day.
Put the two rounds side by side and the trajectory is the headline:
- Seed (June 2026): $200M at a $1B valuation, co-led by Andreessen Horowitz and Kleiner Perkins, with NVIDIA participating.
- Series A (November 2026): $1B, co-led by Kleiner Perkins and Clearlake Capital, with Google joining.
- Elapsed time: under five months.
- Public product: still none; Lor is invite-only.
Kleiner Perkins is the through-line, leading both rounds. When your seed investor comes back to lead a round 5x the size before you’ve launched, that’s a signal the insiders like what they’re seeing on the inside.
⚠️ The Catch
The flashing warning light is the valuation gap. Mirendil did not disclose a Series A valuation, and reports that it was in talks at a $5 billion figure remain unconfirmed by the company (Tech Startups). So treat any “5x in five months” framing as directional, not gospel.
The deeper risk is that all three proof points are things Mirendil built for itself. A coding model it trained, a context window it measured, a beta it runs internally. None of that is the same as outside customers, on outside problems, paying outside money. The company says it wants universities and other organisations to run the platform on their own work, but that’s a plan, not a revenue line.
And “AI that improves AI” is the most crowded, most hyped pitch in the industry right now. A billion dollars buys a lot of compute and a lot of hiring, but it also buys a lot of expectation. If Lor’s agents turn out to be good demos rather than dependable researchers, the markdown will be brutal.
🎯 What Happens Next
Watch for three things. First, when Lor opens up beyond private beta, and whether the “hundreds of agents for days” claim holds when strangers are driving. Second, the first named external customer, because a university or lab willing to go on record is worth more than any internal benchmark. Third, the hiring sheet, since a billion dollars with a two-person founding story means the next twenty hires define whether this is a lab or a legend.
Expect rivals to sharpen their own “automated researcher” messaging within weeks. When money moves this fast, positioning follows.
🧩 Bigger Picture
Mirendil is a clean read on where 2026 venture capital is pointing: not at applications, but at the machinery that makes the models. The same week Mirendil closed its billion, Q-CTRL raised $290 million for quantum control software and LangChain reportedly raised $300 million for agent tooling (Tech Startups). The capital is flowing to the picks and shovels one layer below the chatbots.
It’s also a story about talent gravity. Two researchers walked out of a frontier lab and, within a year, commanded $1.2 billion on the belief that the lab’s methods could be bottled and sold. Whether that’s the democratisation of AI research or just the latest place for a lot of money to wait, the next twelve months will tell.
For now, Mirendil has the rarest thing in the business: a billion dollars and nothing it’s required to show anyone yet.
Sources
- Mirendil — Series A announcement
- Yahoo Finance / PR Newswire — Mirendil raises $1B
- Tech Startups — Startup Funding News, 13 Nov 2026
- The Next Web — Mirendil $200M seed
- TechFundingNews — ex-Anthropic researchers raise $200M
❓ Quick answers
How much did Mirendil raise in its Series A?
Mirendil raised a $1 billion Series A announced on 12 November 2026, co-led by Kleiner Perkins and Clearlake Capital (source: Mirendil).
Who founded Mirendil?
Behnam Neyshabur (ex-Alphabet/Gemini reasoning, then Anthropic) and Harsh Mehta (who built Anthropic's internal AI-research platform), who left Anthropic in December 2025 and founded Mirendil in early 2026.
What does Mirendil build?
Models and agents that perform AI-research tasks like designing experiments, tuning hyperparameters and evaluating models. Its beta product Lor runs hundreds of agents for days at a time.





