US frontier lab Mirendil, founded by ex-Anthropic researchers, raises a US$200m seed at ~US$1bn
Andreessen Horowitz and Kleiner Perkins co-lead with Nvidia; a ~20-person team from Anthropic, xAI, DeepMind and OpenAI is building AI that automates AI research
Add to a list
No lists yet.
Summary
Mirendil, a San Francisco frontier lab founded by researchers who left Anthropic in late 2025, raised a US$200m seed at roughly a US$1bn valuation, one of the largest AI seed rounds on record, co-led by Andreessen Horowitz and Kleiner Perkins with Nvidia participating. Founders Behnam Neyshabur, previously years on Alphabet's Gemini reasoning work, and Harsh Mehta are building models that automate AI research itself, running autonomous coding-agent loops over their own GPUs to design experiments with decreasing human input. The roughly 20-person team is drawn from Anthropic, xAI, Google DeepMind and OpenAI. Reported publish dates range from 26 to 29 June 2026.
The split
US tech-culture coverage seizes on the safety framing, a startup openly building self-improving AI and telling the public not to fear it. European venture press reads the same round as evidence of seed-stage inflation, US$200m before a product, and of talent churn out of the big labs. Both note the pitch that frontier research should not be confined to a few incumbents, the argument now used to justify each new well-funded splinter lab.
By the numbers
- US$200m, seed round, among the largest on record.
- ~US$1bn, seed-stage valuation.
- ~20, team size, from Anthropic, xAI, DeepMind and OpenAI.
- a16z and Kleiner Perkins co-leads, with Nvidia.
- Reported June 26 to 29, 2026.
Why it matters
Billion-dollar seed valuations for pre-product labs building recursive AI research show capital chasing the "AI that improves AI" thesis at the earliest, riskiest stage, and accelerating the talent exodus from Anthropic and OpenAI into funded splinters. It also puts self-improving-AI safety questions into a commercial venture, not just a research debate.
What to watch
- Whether Mirendil publishes anything demonstrating recursive self-improvement.
- Continued senior departures from the big labs into new startups.
- How safety scrutiny treats a company explicitly automating AI R&D.