Mirendil, a startup founded in early 2026 by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, is in talks to raise up to $1 billion in new funding at a roughly $5 billion valuation. Kleiner Perkins is reportedly considering leading the round, with Andreessen Horowitz in discussions to participate — both firms also led Mirendil's earlier $200 million seed round at a $1 billion valuation just months ago.
The company, staffed by roughly 20 researchers pulled from Anthropic, xAI, Google DeepMind and OpenAI, is building AI systems aimed at automating AI research itself — a technique known as recursive self-improvement, where models help train and refine their own successors. Mirendil has not shipped a public product yet and reportedly plans to launch a frontier model for engineering and research work by early 2027, relying on Google Cloud TPUs and Nvidia infrastructure.
The fivefold valuation jump in a matter of months underscores how aggressively investors are chasing labs working on self-improving AI, even before any product has reached the market. Safety remains the central open question: building strong safeguards around systems designed to improve themselves, while still expanding access to the technology beyond the largest labs.