Mirendil, founded in early 2026 by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, announced a $200 million seed round at a $1 billion valuation on June 24, 2026. The round was co-led by Andreessen Horowitz and Kleiner Perkins, with Nvidia as a participating investor. The team includes roughly 20 engineers and researchers from Anthropic, xAI, Google DeepMind, and OpenAI.
Mirendil is building frontier AI models specialized in automating AI research and development tasks—experimental design, hyperparameter search, model evaluation, and iterative training loops. The company packages these capabilities as a platform intended to help scientists, engineers, and research institutions build specialized models without frontier-lab-scale resources. The first model and product launch is planned for the coming months.
Mirendil's founders argue this is an engineering problem, not a safety barrier. Harsh Mehta built the first version of Anthropic's autoresearch platform as a team of one, then scaled it within the company. Neyshabur spent 7+ years on AI-for-science at Google and Anthropic. Both are betting that the next structural shift in AI is democratizing research capability, not consumer capability.
For architects: this is the frontier lab exit play applied to research infrastructure itself. When leading research organizations (Anthropic, xAI, DeepMind) start building internal AI-for-AI-research systems, ex-employees spinning out to commercialize that pattern signals mature product-market fit in one domain and appetite from top-tier VCs (a16z + Kleiner Perkins + Nvidia) to fund the standalone version. If it works, recursive self-improvement moves from labs to platforms.