Mirendil Series A: ex-Anthropic researchers raise $1 billion for AI lab agents

Representative image: Tho-Ge / Pixabay
Mirendil, founded by former Anthropic researchers, raised a $1 billion Series A led by Kleiner Perkins and Clearlake on 12 November 2026.
What Happened
Mirendil, a San Francisco artificial intelligence start-up set up by former Anthropic researchers, announced on Thursday, 12 November 2026 that it has raised $1 billion in a Series A round led by Kleiner Perkins and Clearlake Capital. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also joined the round, according to the company's announcement. Mirendil did not disclose a valuation. The company is building Lor, a platform in which large numbers of AI agents write code, run experiments and work through results with little human input, a job normally done by teams of researchers inside an AI lab.
Mirendil: key facts
- Amount: $1 billion Series A, announced on 12 November 2026.
- Lead investors: Kleiner Perkins and Clearlake Capital; Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also took part.
- Valuation: not disclosed by the company. Earlier press reports had said it was in talks to raise at about $5 billion.
- Founders: Behnam Neyshabur (chief executive) and Harsh Mehta (chief technology officer), both formerly researchers at Anthropic.
- Earlier funding: Mirendil launched in June 2026 with a seed round of about $200 million in which Andreessen Horowitz and Kleiner Perkins took part.
- Product: Lor, introduced on 28 October 2026 and still in an invitation-only private beta, free during the beta.
- Company claim: one researcher working with its agents trained a base model, Mirendil-Base, from scratch in 3.5 months with less than $10 million of computing, a figure the company has not had independently checked.
Mirendil: why it matters
Most AI start-ups that raise money at this scale are building a model or a consumer product. Mirendil is selling something different: a way to automate the research process itself. In Lor, the company says, hundreds of software agents can run at once for days or weeks, writing code, starting training runs, renting computing power and checking what worked. An agent here means an AI program that can take a series of actions towards a goal on its own, rather than answering one question at a time. If that approach works as described, a small team could do the kind of work that today needs a large lab and a large budget.
That is why the claim about Mirendil-Base attracts attention. A base model is the first, general version of a language model before it is tuned for a particular use. Training one from scratch normally takes a team and a large computing bill. Mirendil says a single researcher did it with its agents in about three and a half months for under $10 million. This is the company's own figure; the round's announcement did not include outside testing, and readers should treat it as a claim until others can check it.
The round also shows how much money is still flowing to small teams with a strong research pedigree. Investors are backing founders who came out of the big labs, betting that tools for making AI will be as valuable as the models. Google's place on the investor list is notable because Mirendil had already signed a multiyear cloud deal with Google in August to use its TPU chips (Google's own AI processors) and Nvidia hardware for training and running models.
For Indian founders and engineers, the story is a signal of where global capital is moving. The value is shifting from building yet another chatbot to building the systems that make research cheaper and faster. Indian AI teams working on agents, developer tools and model training efficiency sit in the same space, and a lower cost of training models matters most to countries and companies that cannot spend billions on computing.
Peepals calculation: The new round is about five times the seed round raised in June ($1,000 million divided by $200 million = 5), and it came roughly five months later. Measured against the claimed training cost, the round is more than 100 times what Mirendil says one base model cost to train ($1,000 million divided by under $10 million = over 100). Our calculation from the reported figures.
| Round | Date | Amount | Led or joined by |
|---|---|---|---|
| Seed | June 2026 | About $200 million | Andreessen Horowitz, Kleiner Perkins |
| Series A | 12 November 2026 | $1 billion | Kleiner Perkins, Clearlake Capital (lead) |
Impact
Short-term: Mirendil now has the money to widen access to Lor beyond its private beta and to buy the computing power its agents need. The company has not said when Lor will open to all users or what it will cost after the free beta.
Long-term: If agent-run research proves reliable, the cost and time of building new AI models could fall, which would let smaller labs, universities and start-ups, including those in India, compete with the largest firms. If the claims do not hold up in wider use, the round will be read as another sign of very high valuations in AI.
Who is affected: AI researchers and engineers whose daily work Lor aims to automate; start-ups building agent tools, which now face a well-funded rival; cloud providers selling computing time; and investors weighing whether research automation is a durable business.
What to watch next: whether Mirendil names a date for opening Lor beyond its invitation-only private beta, and whether any outside group tests its Mirendil-Base training claim.
Key Takeaway
Mirendil's $1 billion Series A, five times its June seed round, is a large bet that AI agents can do much of the work of an AI research lab.
Questions and Answers
How much did Mirendil raise in its Series A and who led it?
Mirendil raised $1 billion in a Series A announced on 12 November 2026, led by Kleiner Perkins and Clearlake Capital, with Google, Battery Ventures and others also investing.
What is Mirendil's Lor platform?
Lor is Mirendil's platform in which many AI agents write code, run experiments and review results for days at a time. It was introduced on 28 October 2026 and is in an invitation-only private beta.
Who founded Mirendil?
Mirendil was founded by Behnam Neyshabur, its chief executive, and Harsh Mehta, its chief technology officer, both former researchers at Anthropic.
Has Mirendil's claim about training a model for under $10 million been verified?
No. The claim that one researcher used its agents to train Mirendil-Base in 3.5 months for under $10 million of computing comes from the company and has not been independently confirmed.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
By Elena D’Souza, Startups & Innovation CorrespondentCovers Start-ups & Opportunities and more for Peepals Global. All stories by Elena D’Souza
SAFF Championship final: India face Pakistan in Dhaka on November 17
Estonia to close Russia border crossings after attack on drone maker Threod
Get the day's stories in your inbox
Peepals Weekly: the week decoded, free. One email a week, unsubscribe in one click.
Elena D’Souza is a pen name of the Peepals Global Editorial Team, which researches, fact-checks and publishes every story. The portrait is an illustration.
Responsible for this story: the Peepals Global Editorial Team, led by editor Prashant Sonar. How we report









