Mirendil, a San Francisco-based startup founded by former Anthropic researchers, is in talks for a new funding round that would value the company at $5 billion. The company is seeking to raise as much as $1 billion in the round, with Kleiner Perkins reportedly set to lead the investment and venture giant Andreessen Horowitz (a16z) also participating.

The pace of the ascent is striking: roughly three months ago, Mirendil closed a $200 million seed round at a $1 billion valuation. The new round would represent a fivefold increase in valuation within a single quarter.

Founded by Anthropic Alumni

Mirendil was founded in December 2025 by Behnam Neyshabur and Harsh Mehta, both of whom previously worked as researchers at Anthropic. Neyshabur co-led the lab's scientific AI reasoning team, while Mehta served as a senior research scientist. The founding team numbers around 20 specialists drawn from top-tier AI organizations.

The company's earlier seed round, completed in June 2026, was co-led by a16z and Kleiner Perkins, with chipmaker NVIDIA also participating. The return of both co-leads within roughly a quarter — scaling from a $200 million seed to a round of up to $1 billion — gives the new raise continuity: the firms that set Mirendil's $1 billion price are now, per reports, positioned to underwrite it at $5 billion.

What Mirendil Is Building

At the core of the startup's work is the development of self-improving AI systems designed to automate complex AI research and development tasks. In effect, Mirendil is building AI that can do the work of AI researchers: designing experiments, iterating on model architectures, and pushing the boundaries of what is possible without requiring a team of hundreds of PhD holders.

The company frames this mission as an effort to democratize frontier-level AI research capabilities. Today, only a handful of organizations — among them OpenAI, Google DeepMind, Anthropic, and Meta — have the resources to conduct cutting-edge AI research. That concentration of capital and talent is precisely the constraint Mirendil's approach is built to address: its bet is that automated research systems, rather than ever-larger human teams, can widen access to the frontier.

Compute Partnership

Training self-improving models requires enormous amounts of computing power. To meet those demands, Mirendil secured a $100 million Google Cloud deal in August 2026, giving the startup access to TPUs, NVIDIA GPUs, and managed clusters. The size of that commitment is a yardstick for the capital intensity of the field: at this scale of research, compute access is a core input rather than an add-on, and large cloud deals have become the standard route for startups to secure it.

If the $5 billion round closes, Mirendil will have raised approximately $1.2 billion in total funding within roughly nine months of its founding. Since the round remains in the talks stage, whether the reported size and valuation hold through to a signed deal is the immediate question — the last datapoint needed to confirm the fivefold, one-quarter jump.