Mirendil, a startup built by former Anthropic researchers, is discussing a new funding round that would value the company at $5 billion. Bloomberg broke the story, attributing the figure to people with direct knowledge of the negotiations. The talks are ongoing and no deal has closed, so the number is a proposed price, not a completed round.
Kleiner Perkins is negotiating to lead the round, which Bloomberg’s sources say could bring in as much as $1 billion in fresh capital. Andreessen Horowitz is weighing whether to join the round as well. Both firms backed Mirendil’s $200 million seed round three months ago, which valued the company at $1 billion, meaning the price under discussion now is five times higher in a single quarter.
The technical bet behind Mirendil is recursive self-improvement: building models that can improve themselves with minimal human input. Founded this year, the company has already grown past 20 employees. On its website, Mirendil frames the ambition broadly, saying that “any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab,” and that it wants to make that capability more widely available. That is the company’s own framing of its mission, not an independent assessment.
The new money would go toward training Mirendil’s own frontier model, which Bloomberg’s sources say the company plans to ship for engineering and research use by early next year. The pitch mirrors an internal metric two larger labs already cite: Anthropic says Claude now handles more than a quarter of its own AI research and development, and OpenAI uses Codex the same way inside its own model work. Investors betting on Mirendil are effectively betting that a small team can replicate that self-improvement loop from scratch.
Investors have a name for this category: neo-labs, startups chasing ambitious research agendas well ahead of any near-term product revenue. Mirendil is one of several such bets drawing capital right now. Mira Murati, OpenAI’s former chief technology officer, now runs Thinking Machines Lab, which is negotiating to raise upward of $1 billion at a valuation above $40 billion. That figure comes from a PYMNTS report sourced to The Information. A third entrant, Periodic Labs, focuses on scientific discovery. Liam Fedus, a researcher who previously worked at OpenAI, built the company, and its valuation sits near $1.3 billion.
The math behind the Mirendil talks puts the proposed valuation at 25 times the capital it has raised to date, a multiple in the same range as Thinking Machines Lab’s own funding arithmetic. Neither company has meaningful product revenue to point to yet. What both have is a research thesis and a founding team with frontier-lab pedigree, and investors are pricing that pedigree rather than a P&L.
For operators watching the funding market, the read is this: a five times valuation jump in one quarter, on talks that have not closed, is a signal that venture capital is now pricing recursive self-improvement research as its own asset class, separate from whether any of these teams has shipped a product yet.
Based on PYMNTS reporting (September 22, 2026), which cited Bloomberg’s reporting on Mirendil’s funding talks.
