Mirendil raises $1B Series A for AI research agents in 2026

Lead

Mirendil, a San Francisco AI lab founded by former Anthropic researchers, announced a $1 billion Series A on November 12, 2026. Kleiner Perkins and Clearlake Capital led the round. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also participated, alongside other new and existing investors. The company did not disclose the valuation.

The raise arrives five months after a $200 million seed round co-led by Andreessen Horowitz and Kleiner Perkins at a $1 billion valuation. Capital raised has grown fivefold in that span.

Who Is Behind Mirendil?

Mirendil was founded by CEO Behnam Neyshabur and Harsh Mehta, who met at Google in 2019. Neyshabur spent more than five years at Alphabet, where he led Gemini reasoning research. Mehta built Anthropic's internal AI-research platform. Both joined Anthropic in late 2024 and left in December 2025.

The company launched with about 20 researchers drawn from Anthropic, xAI, Google DeepMind and OpenAI. Its announcement describes it as "still a small team," which is the point of the pitch: when agents do the hands-on work, headcount stops capping what a lab can build.

What Does Mirendil Actually Build?

Mirendil builds models and an agent platform that automate the routine labor of research: writing code, running experiments and provisioning compute. The company frames this as opening frontier research to scientists and engineers outside the largest labs.

Its products fall into three groups:

• Lor, a cloud platform in private beta, introduced on October 28, 2026. It orchestrates hundreds of agents that write code and run experiments over extended periods, on machines the user selects. It works with Claude Opus, GPT-6 Sol and open-source models.
• Fern, a model with a context window of up to 10 million tokens, roughly 500,000 lines of code.
• Mirendil-Base, a competitive coding model.

The framing matters. The target is AI research itself, with the broader scientific-research use case a stated extension rather than a shipped product. No customer names or revenue figures have been disclosed.

How Credible Is the Efficiency Claim?

The claim is Mirendil's own and has not been independently verified. In an October 21, 2026 disclosure, the company said one researcher working with agents pretrained Mirendil-Base from scratch in 3.5 months, using less than $10 million in compute. Agents wrote more than 99% of the code, according to the company, which also said it used only public data, with no distillation and no licensed datasets.

If the figures hold, they undercut the assumption that frontier-adjacent models require nine-figure training budgets and hundreds of engineers. They also leave open how Mirendil-Base performs against models from labs that spent far more. A coding model trained cheaply is a proof of method, not a proof of competitiveness.

Why Did the Valuation Jump So Fast?

The valuation question is partly unresolved. A September Bloomberg report put the target at about $5 billion, roughly 25 times the capital raised at that point, while the final Series A figure remains undisclosed. At a $5 billion mark, the company would be valued at 25 times its seed funding with fewer than 30 employees and a product still in private beta.

That is the skeptical read of the round. Investors are paying for a team with Alphabet and Anthropic pedigree and for a thesis on recursive self-improvement, in which AI systems accelerate the development of better AI. Revenue is not part of the case yet. Google's participation adds a former employer of both founders to the cap table, which signals competitors and suppliers are hedging rather than ignoring the approach.

What Comes Next for Mirendil?

Mirendil plans to launch a frontier model for engineering and research work by early 2027, according to earlier reports on its fundraising. Proceeds will go toward compute and hiring, though the company has not published a split. The next test is whether Lor moves from private beta to paying users and whether outside researchers can reproduce the cost savings the company claims.

Three scenarios stand out. If the efficiency claims replicate, larger labs will face pressure to adopt similar agent workflows, and the cost of entering model development falls. If they do not, the $1 billion becomes a large bet on a method that works only inside Mirendil. A third path is acquisition interest from a hyperscaler, given Google's seat in the round, though no such talks have been disclosed.

Outlook

Mirendil now holds $1.2 billion in total funding across two rounds, a team of roughly two dozen people, and one unverified efficiency benchmark. The round tests whether agent-driven research labs can scale without scaling headcount. Evidence arrives with the early-2027 model release and any independent evaluation of Lor.