AI startup Mirendil has emerged from stealth with a $200 million seed round, pushing its valuation to $1 billion and earning unicorn status. The round was led by Andreessen Horowitz (a16z) and Kleiner Perkins, with participation from Nvidia. The company was founded by Behnam Neyshabur and Harsh Mehta, both former researchers at Anthropic who left shortly after the December 2025 release of Claude Opus 4.5.
Two Anthropic Veterans Strike Out on Their Own
Behnam Neyshabur serves as CEO, having previously led the scientific AI reasoning team at Anthropic. Co-founder Harsh Mehta also came from Anthropic's research division. The pair first met while working at Google in 2019, joined Anthropic together in late 2024, and left within a year after the Claude Opus 4.5 launch. The founding team also includes early xAI member Shayan Salehian and MIT graduate Tara Rezaei. Mirendil currently operates out of downtown San Francisco with about 20 technical staff.
Building AI That Rewrites Itself
Mirendil's core technology is recursive self-improvement — AI that can autonomously rewrite, train, and upgrade itself without human engineers manually intervening each time. The company isn't positioning itself to compete head-on with OpenAI or Anthropic. Instead, it aims to create an "AI that accelerates AI research," enabling scientific labs and enterprises to build and control their own specialized models without relying on a handful of frontier labs. Target applications include medical research and materials science.
Is Self-Improvement Feasible?
There's a fine line here. Anthropic disclosed that as of May 2026, Claude had written over 80% of the company's code — frontier labs already use AI to speed up model development. Yet those same labs prohibit external developers from using their large models to train competing products, and have quietly restricted answers to AI-development-related queries without notifying users. Mirendil takes a direct stance: recursive self-improvement is the fastest path to accelerating science, and it can be done under safety oversight. This puts Mirendil at odds with frontier labs' caution. Many AI safety researchers warn that self-upgrading models pose unpredictable risks; Mirendil frames the technology as a tool for accelerating science, arguing that safety regulation can run in parallel. The startup plans to release models and products in the coming months to gather early user feedback.