Mirendil, a startup founded by two former Anthropic researchers, has raised $200 million in seed funding at a $1 billion valuation, backed by Andreessen Horowitz (a16z), Kleiner Perkins, and Nvidia. The 20-person company is building infrastructure that enables artificial intelligence systems to conduct their own research and development, a capability that has historically been locked behind the walls of major tech companies.

The founding team brings deep credentials from the AI frontier. CEO Behnam Neyshabur previously led Anthropic's Discovery team, which focused on enabling AI to perform long-horizon reasoning tasks similar to how scientists and engineers work. CTO Harsh Mehta built Anthropic's first automated AI research platform. Both left Anthropic in December 2025, shortly after the company released Claude Opus 4.5, its next-generation model with powerful agentic capabilities.

Why Are Tech Giants Hoarding AI Research Capabilities?

The market opportunity Mirendil is targeting stems from a structural imbalance in the AI industry. According to Anthropic's public disclosures, over 80 percent of the code merged into the company's codebase was generated by Claude as of May 2026, with the typical engineer merging eight times more code per day in the second quarter of 2026 than in 2024. Similarly, Google's AlphaEvolve system is already being used to optimize data center scheduling, chip design, and AI training pipelines. Yet these giants explicitly restrict external developers in their user agreements from using their models to train competing products.

"This is merely the normal reaction of tech giants acting as rational economic actors," remarked Matt Bornstein, an a16z investor, noting that the restriction has created a structural void in the market that an independent company must fill.

Matt Bornstein, Investor at Andreessen Horowitz

Mirendil's positioning is not to directly conduct drug discovery or materials design for scientists, but to provide research institutions lacking top-tier AI research and development capabilities with the tools to build their own. A drug development company may deeply understand disease mechanisms and possess laboratory resources, but lack the ability to train state-of-the-art AI models. Assembling an AI team, building training infrastructure, processing data, and connecting to real-world experiments is prohibitively expensive and inefficient.

How Does Recursive Self-Improvement Work in Scientific Research?

The technological approach Mirendil has chosen is called "recursive self-improvement," where AI systems continuously optimize their own code and models, eliminating dependence on manual parameter tuning by human engineers. Neyshabur calls it the shortest path to "AI-accelerated science" and argues it can be pursued safely under human supervision.

The vision sees AI evolving from executing human-set objectives to independently designing experimental pathways and ultimately autonomously proposing scientific hypotheses, training models, and evaluating results. This would usher scientific discovery into a form of automation that could dramatically accelerate breakthroughs across multiple fields.

Steps to Understanding Mirendil's Market Opportunity

• The Capability Gap: Major AI labs generate over 80 percent of their internal code via AI, but refuse to let external developers use their models for competitive products, creating a market void for independent infrastructure.
• The Target Market: Pharmaceutical companies, materials science researchers, robotics firms, and other institutions need specialized AI models but lack the resources to build them from scratch or hire elite AI teams.
• The Solution: Mirendil aims to unbundle frontier AI research and development capabilities from major companies and turn them into a platform where scientists can train and iterate their own specialized models as easily as accessing a utility.

Despite the grand vision, Mirendil has yet to release any product. The company plans to launch its first model and product within the coming months and gather early user feedback. Beyond the two co-founders, the core team includes Shayan Salehian, an early member from Elon Musk's xAI, and MIT graduate Tara Rezaei.

The $200 million seed round ranks among the largest for AI startups in recent years. For context, it trails only Thinking Machines Lab's $2 billion, Safe Superintelligence's $1 billion, and Advanced Machine Intelligence's $1.03 billion, placing it in the same league as Periodic Labs' $300 million. This funding level underscores venture capital's voracious appetite for top-tier AI talent spilling out of major labs.

The company name "Mirendil" is drawn from the Elvish language in Tolkien's Lord of the Rings, meaning "friend of precious things" or "friend of discovery," a fitting moniker for its positioning as new infrastructure for scientific discovery.