Anthropic Alumni Startup Mirendil Raises $1 Billion Series A on Agent-Driven Model Claims

11-month-old San Francisco lab Mirendil announced a $1 billion Series A on November 12, led by Kleiner Perkins and Clearlake Capital, with participation from Google, Battery Ventures, Doerr Capital, Gil & Co, M13, Pear VC and S32 [4]. The round follows a roughly $200 million seed in June 2026 [1], led by Andreessen Horowitz [5]. Mirendil has not disclosed a valuation; press reports from September put the target at about $5 billion, roughly five times the reported seed valuation [1].

Founders and Premise

Mirendil was founded in late 2025 by Behnam Neyshabur (CEO) and Harsh Mehta (CTO), who met at Google in 2019 and later worked together at Anthropic [1]. Neyshabur co-led reasoning research for Gemini at Google and a research team at Anthropic; Mehta worked on Gemini pretraining and led automated AI research work at Anthropic [1]. The team includes early members Shayan Salehian (ex‑xAI) and Tara Rezaei (MIT graduate, former OpenAI student researcher) [1].

The company’s pitch: an agent platform called Lor that automates AI research itself—designing experiments, writing training code, running evaluations and iterating—so a small team can build frontier models at a fraction of the usual labor and compute cost [4][1]. Andreessen Horowitz described the approach as “vibe research” [5].

What Has Been Shipped

Mirendil has released three artifacts, all with performance claims coming solely from the company:

• Mirendil‑Base: a pretrained coding model of 100–500 billion parameters, published October 21 [1]. The company says one researcher, working with agents, pretrained it from scratch in 3.5 months for under $10 million in compute, with agents writing more than 99% of the code [1].
• Fern: a post‑trained version of Mirendil‑Base with a claimed 10‑million‑token context window (about half a million lines of code), released in early access [1][2]. Only a basic long‑context test has been published [1].
• Lor: the agent platform that runs hundreds of research agents writing code and running experiments for days at a time. It entered private beta October 28 [1]. Mirendil says it “runs on Lor every day” [4].

Two larger models, Mirendil‑Base‑Large (0.5–2 trillion parameters) and Mirendil‑Base‑XL (2–10 trillion), were reported as 80% and 19% through pretraining respectively as of October 21 [1].

Benchmark Claims and Gaps

Mirendil compared Mirendil‑Base against several open models on held‑out code using “bits per byte,” a measure of next‑token prediction accuracy (lower is better) [1]. The comparison is not like‑for‑like because some reference models are post‑trained, and bits per byte is a narrow metric that says little about reasoning or real coding tasks [1]. Mirendil also states it does not think it has reached fully autonomous recursive self‑improvement yet [1]. No independent evaluations of Mirendil‑Base, Fern or Lor have been published.

Investor Rationale and Concentration

Kleiner Perkins, which co‑led the seed, doubled down with a Series A co‑lead. Its June investment note summarized the thesis: “Better models do better research. Better research produces better models. The loop is the product” [1]. Clearlake Capital, a private‑equity firm that closed an $14.8 billion flagship fund in June focused on “AI‑driven change,” is an unusual Series A lead [1]. Google participates as both investor and compute supplier under a multiyear Cloud deal worth over $100 million signed in August [1].

Andreessen Horowitz, which led the seed, does not appear on the published Series A investor list [1]. The press release notes “other new and existing investors” without naming them [4].

Revenue, Headcount and Market Context

Mirendil describes itself as “still a small team with a singular focus” [4]; earlier reports put headcount at about 20 researchers and engineers [1]. The company has not disclosed any revenue [1]. The $1 billion Series A for an 11‑month‑old lab with ~20 people and no public revenue data fits a 2026 pattern: Crunchbase News counted 27 venture rounds of $1 billion or more in Q3 alone, accounting for roughly a third of the quarter’s $159 billion in global venture funding [1].

Risks and Unknowns

• All performance claims are company‑run and unverified by outside parties [1].
• The reported ~$5 billion valuation is unconfirmed by Mirendil [1][4].
• Lor remains in private beta; Fern is in early access; the two largest models are still training [1].
• Google is both investor and compute provider, creating a concentration point [1].
• Larger labs are automating their own research with far more data, talent and compute [1].
• AI funding sentiment can shift quickly; Mirendil’s next valuation step will likely depend on outside evaluations and any revenue signal [1].

What Comes Next

Mirendil says the Series A “lets us go further” toward making frontier AI R&D accessible [4]. The near‑term milestones are the completion of Mirendil‑Base‑Large and ‑XL, opening Lor beyond private beta, and—critically—third‑party testing of its models and agent platform. Until those arrive, the $1 billion round and reported $5 billion valuation rest primarily on the founders’ track records and internal benchmarks.