Companies used machine learning for scoring, routing, and forecasts long before November 2022. ChatGPT is when a frontier model sat in a browser that anyone could open. Headlines after that clustered around Anthropic and OpenAI. Job posts and LinkedIn announcements from both shops show how fast those two benches grew. That still leaves Mistral, Cohere, Google Gemini in several cuts, Amazon Nova, and the Microsoft MAI line still coming. I wrote earlier this month about Cloudera putting Mistral weights on customer data. Those names do not vanish because two labs own the news cycle.
Natasha Mascarenhas and Shirin Ghaffary at Bloomberg reported on September 22 that Mirendil is in talks for a new round at a $5 billion valuation, including the investment. Kleiner Perkins is said to lead. Andreessen Horowitz is in the conversation. The talks are not a closed deal. Three months earlier, on June 24, the same two firms led a $200 million seed that valued the company at $1 billion. Nvidia put money in that seed. TechCrunch later reported a multiyear Google Cloud contract above $100 million for tensor processing units and Nvidia graphics processors (Mascarenhas and Ghaffary; Bellan).
The lab wants other labs to run their own research loop
Chief executive Behnam Neyshabur and chief technology officer Harsh Mehta left Anthropic in late 2025 after about a year there. They met at Google in 2019. Neyshabur co-invented the Sharpness-Aware Minimization optimizer and later co-led Anthropic's Discovery work. Mehta built the first version of Anthropic's internal automated research tools, at times as a team of one. Co-founders Shayan Salehian, from the early xAI bench, and Tara Rezaei, a 23-year-old MIT graduate who interned at OpenAI, round out the founding group. The company says the wider team of about 20 came from Anthropic, xAI, Google DeepMind, and OpenAI (Bornstein and Aubakirova; Mirendil).
The pitch is a platform that designs experiments, picks training settings, scores the result, and starts the next run with less human steering each cycle. Neyshabur has described the buyer as a university biology group that needs a drug-target model and does not keep a machine-learning staff. He has used an Alzheimer's risk model as the kind of artifact that group might produce. Andreessen Horowitz called the idea vibe research. Bloomberg's sources said the first frontier model for engineering and research work is aimed at early 2027. Nothing public is in your stack today (The Next Web; PYMNTS).
A leaner shop can still force a price check
A fivefold mark-up in a quarter, with no product page you can trial, tells you investors are paying for the research loop and the people who already ran a version of it inside Anthropic. It also tells you the two largest consumer-facing labs do not lock the category. Newer companies can stay smaller, spend less on brand, and still bid for the same researchers. That is the part that should sit on a vendor review, not the Elvish company name. Mirendil means friend of precious things in Tolkien's language. Cute. The procurement question is whether you need their loop or a smaller model that already sits behind your firewall.
Plenty of work that now hits Claude or GPT can run on Mistral, Nova, or an older Gemini cut. Teams route the hard ticket to the largest model because the seat is already paid and the chat window is open. That habit inflates both the bill and the blast radius. If a system can keep improving itself, the next question is what happens when someone types a goal with no guardrail. Earn this much money. Find a way. I am not writing a collapse story. I am saying a model that can act, browse, and pay for tools can take a path you did not approve. The blunt control, if you ever need it, is still physical. Pull the plug. Keep the box off the public internet.
Dario Amodei published We Must Pace the Frontier on September 12 and asked labs to slow capability gains so evaluators can sit inside the building. Mirendil is hiring and raising toward a self-improving loop in the same month. Both statements can be true at once. One lab wants outside desks. Another lab wants the research cycle to run with fewer humans. You still have to decide which model class a given job is allowed to touch.
Ask product and data science which current jobs need a frontier model and which already run well on Mistral, Nova, Gemini, or an older in-house checkpoint.
Ask security whether any agent with tools and a browser can accept an open-ended money or access goal without a human in the loop.
Treat the Mirendil talks as a reminder that the supplier list will keep growing. Do not wait for a $5 billion seed story to inventory which model class each workflow is allowed to use.
Mascarenhas, Natasha, and Shirin Ghaffary. "Ex-Anthropic Staffers' Self-Improving AI Startup in Talks to Raise at $5 Billion Value." Bloomberg, 22 Sep. 2026, https://www.bloomberg.com/news/articles/2026-09-22/ex-anthropic-staffers-ai-startup-in-talks-to-raise-at-5-billion-value.
Bellan, Rebecca. "Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI." TechCrunch, 6 Aug. 2026, https://techcrunch.com/2026/08/06/exclusive-mirendil-inks-100m-google-cloud-deal-to-scale-self-improving-ai/.
Bornstein, Matt, and Malika Aubakirova. "Investing in Mirendil." Andreessen Horowitz, 24 June 2026, https://a16z.com/announcement/investing-in-mirendil/.
Mirendil. "Announcing Mirendil." 24 June 2026, https://mirendil.com/news/announcing-mirendil/.
"Ex-Anthropic researchers raise $200M for self-improving AI." The Next Web, 29 June 2026, https://thenextweb.com/news/mirendil-200m-seed-ai-research.
"Anthropic Alumni Startup Sees Valuation Jump Fivefold." PYMNTS, 22 Sep. 2026, https://www.pymnts.com/news/artificial-intelligence/2026/anthropic-alumni-startup-sees-valuation-jump-fivefold/.
Hamid, Mamoon. "Mirendil: Building the system that builds systems." Kleiner Perkins, 24 June 2026, https://www.kleinerperkins.com/perspectives/mirendil-building-the-system-that-builds-systems/.
