The most expensive research assistant in the world doesn't draw a salary, doesn't sleep, and can coordinate hundreds of parallel experiments for weeks at a time. It is an AI agent, and it is the core product of Mirendil, a San Francisco startup that believes the future of artificial intelligence research will be conducted not by sprawling teams, but by individual scientists orchestrating armies of software agents [Perplexity Sonar Pro Brief, Nov 2026].
This vision, backed by an eye watering $1.2 billion in disclosed funding within six months, positions Mirendil at the intersection of two explosive trends: the pursuit of artificial general intelligence and the emerging category of "agentic" AI that can plan and execute complex, long running tasks [Yahoo Finance / PR Newswire, Nov 2026]. The company's premise is that its own technology, which it uses daily, can compress the timeline and cost of frontier AI development, a claim it supports with an early, striking data point. The company reports that a single researcher trained its Mirendil Base coding model from scratch in three and a half months using less than $10 million in compute [Yahoo Finance / PR Newswire, Nov 2026].
The Wedge: From Internal Tool to Commercial Product
Mirendil is commercializing the very system it built to accelerate its own work. Its flagship product, Lor, is an agent platform currently in a private beta. The company describes it as software that can coordinate "hundreds of agents" to write code, run experiments, and provision computing resources autonomously for days or weeks [Perplexity Sonar Pro Brief, Nov 2026]. This is not a chatbot for debugging; it is an automation layer intended to manage the entire iterative cycle of AI research and development.
The company has also developed two AI models that showcase the output of this process. Mirendil Base is a coding model trained from scratch on public data, and Fern is a post trained model notable for its massive, 10 million token context window about half a million lines of code at once [explainx.ai Blog]. These models serve as both proof points and potential components within the broader Lor ecosystem. The commercial bet, however, is on the platform, not just the models. Mirendil is selling the promise of a force multiplier for AI labs, large and small.
A Founding Team Forged in Frontier AI
The ambition is matched by a founding team with deep roots in the labs Mirendil aims to augment. The co founders are alumni of Google DeepMind, Anthropic, xAI, and OpenAI, giving them firsthand experience with the scale and challenges of modern AI research [Perplexity Sonar Pro Brief, Nov 2026].
| Founder | Role | Key Prior Experience |
|---|---|---|
| Behnam Neyshabur | CEO | Co led Discovery team at Anthropic; co led Blueshift team at Google DeepMind; co inventor of SAM algorithm [Perplexity Sonar Pro Brief, Nov 2026]. |
| Harsh Mehta | CTO | Built the first version of Anthropic's internal AI research platform; worked on pretraining at DeepMind [Perplexity Sonar Pro Brief, Nov 2026]. |
| Shayan Salehian | Co Founder | Previously a Member of Technical Staff at xAI [Perplexity Sonar Pro Brief, Nov 2026]. |
| Tara Rezaei | Co Founder | Previously at OpenAI and an MIT graduate [LinkedIn, Tara Rezaei]. |
This concentration of talent is a significant part of the investor story. The team left Anthropic in December 2025 to start Mirendil, bringing with them a specific vision for how AI can be used to build better AI, faster [Perplexity Sonar Pro Brief, Nov 2026].
Capital and Compute at a Staggering Scale
Mirendil's funding trajectory is as accelerated as its research goals. The company raised a $200 million seed round in June 2024, led by Andreessen Horowitz and Kleiner Perkins, at a valuation reported to be roughly $1 billion [Danial Taheri, CFA, Retrieved 2026]. Just five months later, it closed a $1 billion Series A round co led by Kleiner Perkins and Clearlake Capital, with participation from a who's who of tech investors including Google, Nvidia, and Battery Ventures [Yahoo Finance / PR Newswire, Nov 2026].
2026 Seed | 200 | M USD
2026 Series A | 1000 | M USD
This capital is not sitting idle. In August 2024, Mirendil signed a deal worth over $100 million with Google Cloud, securing extensive access to Tensor Processing Units (TPUs), NVIDIA GPUs, and managed training clusters [TechCrunch, Aug 2026]. The partnership provides the raw computational horsepower required to train large scale models and, presumably, to run the intensive agent workloads that Lor manages.
The Risks in Automating Discovery
For all its promise, Mirendil's path is fraught with technical and commercial uncertainty. The field of AI research automation is nascent, and Lor remains in a private beta with no publicly disclosed commercial customers or pricing [Perplexity Sonar Pro Brief, Nov 2026]. The company's own use case is compelling, but product market fit with external AI research teams is unproven. Furthermore, the core technology agentic AI systems that operate reliably over long time horizons is itself a major unsolved research problem. A platform that can flawlessly orchestrate weeks of experimentation is a monumental software engineering challenge.
The competitive landscape is also unclear but potentially formidable. Every major cloud provider and AI lab is investing heavily in internal tooling for research efficiency. Mirendil's answer to these risks appears to be a combination of its team's unique expertise, its early technological demonstrations, and its formidable war chest. The company is hiring for foundational roles in finance, company building, and go to market, signaling a shift from pure research to commercialization [Perplexity Sonar Pro Brief, Nov 2026].
What to Watch in the Next Twelve Months
The coming year will be critical for Mirendil as it transitions from a well funded research project into a commercial entity. The key milestones to watch are the public launch of Lor, the announcement of its first enterprise customers beyond Google Cloud, and any peer reviewed publications validating the efficiency gains of its agentic approach. The company must demonstrate that its tools can generalize beyond its own walls and provide tangible value to other organizations engaged in costly AI R&D.
The disease state Mirendil is ultimately addressing is the prohibitive cost and slow iteration speed of frontier AI research. The patient population is every institution, from startups to national labs, that wants to participate in the race for advanced AI but lacks the resources of a Google or an OpenAI. Today, the standard of care is a labor intensive process involving large, highly specialized teams manually designing experiments, writing training code, and managing compute clusters. It is slow, expensive, and difficult to scale. Mirendil's bet is that the lab of the future will be run by a scientist and an AI co pilot, turning months of work into weeks and democratizing access to the frontier. The next phase of its experiment begins now.
Sources
- [Perplexity Sonar Pro Brief, November 2026] Mirendil company and product overview
- [Yahoo Finance / PR Newswire, November 2026] Mirendil Raises $1B to Democratize Frontier AI R&D | https://uk.finance.yahoo.com/news/mirendil-raises-1b-democratize-frontier-173000747.html
- [explainx.ai Blog] Details on Fern model context window
- [Danial Taheri, CFA, Retrieved 2026] LinkedIn post on Mirendil seed valuation
- [TechCrunch, August 2026] Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self improving AI | https://techcrunch.com/2026/08/06/exclusive-mirendil-inks-100m-google-cloud-deal-to-scale-self-improving-ai/
- [LinkedIn, Tara Rezaei] Tara Rezaei's professional background