Mirendil

Mirendil is an American artificial intelligence lab founded in 2026 by Behnam Neyshabur, who previously led Anthropic's scientific AI reasoning team, and Harsh Mehta, a senior research scientist at the same lab. The company is headquartered in San Francisco and develops AI models intended to accelerate scientific discovery in biology and materials science by generating hypotheses and simulating outcomes. As of March 2026, Mirendil was reported to be raising $175 million at a $1 billion valuation in a round co-led by Andreessen Horowitz and Kleiner Perkins.

At a glance

• Founded: 2026 in San Francisco. Neyshabur and Mehta departed Anthropic in December 2025.
• Status: Private. Reported funding round under negotiation as of March 2026.
• Funding: $175 million targeted at a $1 billion valuation in a round reported as co-led by Andreessen Horowitz and Kleiner Perkins. Round terms not finalized as of April 2026.
• CEO: Behnam Neyshabur (co-founder, formerly led Anthropic's scientific reasoning team and a senior staff researcher at Google DeepMind).
• Other notable leadership: Harsh Mehta (co-founder and chief technology officer, formerly senior research scientist at Anthropic), Shayan Salehian (formerly at xAI), Tara Rezaei (formerly at OpenAI).
• Open weights: None disclosed.
• Flagship products: No public model release as of April 2026. Research focus on AI for biology and materials science.

Origins

Mirendil was founded in early 2026 by Behnam Neyshabur and Harsh Mehta, who departed Anthropic in December 2025 to start the company. The founding cohort grew with the addition of Shayan Salehian, formerly at xAI, and Tara Rezaei, a former OpenAI intern. The company is headquartered in San Francisco.

Neyshabur's career background combines academic and industry research. He earned his PhD in computer science from TTI-Chicago under Nathan Srebro's supervision, completed postdoctoral research at NYU and the Princeton Institute for Advanced Study under Sanjeev Arora and Yann LeCun, and joined Google in 2019 as a senior staff research scientist and Gemini team lead. At Google, he co-led the Blueshift team, focused on improving foundation-model reasoning, with research contributions including the Minerva model (the 2022 inaugural language model demonstrating strong mathematical capability) and adjacent improvements in reasoning across the Gemini and Gemma 2 families. In late 2024 he transitioned to Anthropic as a principal research scientist and co-led the Discovery team, with the goal of building "an AI Scientist/Engineer." Under his leadership, the Discovery team contributed to a reported threefold improvement in computer-use capability of Claude Sonnet 4.5.

Mehta had been a senior research scientist at Anthropic during a comparable period, with research contributions on scientific-reasoning capability and adjacent areas. He joined Neyshabur as Mirendil's chief technology officer when the two departed Anthropic.

The founding announcement of Mirendil was reported by industry publications in March 2026, by which date the company had operated for approximately three months following the December 2025 Anthropic departures. The reported round of $175 million at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins, was characterized publicly as in negotiation rather than closed.

Mirendil is one of several "neo-labs" founded in 2025 to 2026 by senior researcher departures from frontier AI labs, alongside Adaption Labs (founded by ex-Cohere executives), Humans& (founded by ex-Anthropic, ex-xAI, and ex-Google researchers), Elorian (founded by ex-Google DeepMind and ex-Apple researchers), and Periodic Labs (founded by ex-OpenAI scientific-research leaders).

Mission and strategy

Mirendil's stated mission is to use AI to accelerate scientific research in biology and materials science. The company has framed its objective in terms of building AI models that "generate hypotheses and simulate outcomes" to compress the cycle time of scientific discovery from drug development through advanced materials research.

The strategy combines three threads. First, foundational research on AI architectures specialized for scientific reasoning, building on Neyshabur's prior leadership of the scientific-reasoning programs at Anthropic and Google. Second, vertical specialization on biology and materials science as the principal application domains, in contrast to the general-purpose-frontier-model strategies of OpenAI, Anthropic, and Google DeepMind. Third, a research-first commercial posture, with no consumer or developer product surface disclosed and substantial early capital allocated to compute and senior-research-team hiring.

The competitive premise is that vertical specialization in scientific domains produces capability advantages that general-purpose frontier models cannot match within the same compute budget, and that biology and materials science represent particularly attractive targets given the volume of structured data, the verifiability of outcomes through laboratory experimentation, and the commercial potential of accelerated drug discovery and materials engineering.

The competitive premise aligns with the broader 2025-vintage "neo-lab" framing, in which specialized AI labs founded by frontier-lab alumni pursue narrower, domain-specific applications rather than general-purpose models. Mirendil's specific positioning emphasizes scientific reasoning and hypothesis generation as the principal research surface.

Models and products

• Foundational research output. Mirendil's primary public output as of April 2026 consists of the founding-team announcement and the company's stated technical thesis. No model architecture or capability target has been comprehensively disclosed.
• No shipped models, APIs, or open weights. The company has not released a model, an API, or open weights as of April 2026.

The commercial distribution strategy beyond research output has not been publicly stated. Whether Mirendil intends to ship its own laboratory-deployed scientific-reasoning product, license capabilities to pharmaceutical and materials-science partners, or pursue a hybrid posture has not been disclosed.

Benchmarks and standing

Mirendil has not released a model and is not represented on the standardized capability leaderboards as of April 2026. The company's standing rests on the founders' research credentials at Anthropic and Google DeepMind, the reported lead-investor signals from Andreessen Horowitz and Kleiner Perkins, and the targeted $1 billion valuation.

The standardized leaderboards measure general-language and reasoning capability rather than scientific-domain capability, and may not be the appropriate evaluation framework for Mirendil's research output even after a public model release. Domain-specific scientific benchmarks (such as those used for AlphaFold 3 and adjacent biological-AI evaluations) are likely a closer comparator framework.

Leadership

As of April 2026, Mirendil's named leadership consists of:

• Behnam Neyshabur, co-founder and chief executive. Formerly co-led Anthropic's Discovery team focused on scientific AI reasoning. Prior senior staff research scientist and Gemini team lead at Google for over five years. Public face for the company on technical claims and product framing.
• Harsh Mehta, co-founder and chief technology officer. Formerly senior research scientist at Anthropic.
• Shayan Salehian, founding team member. Formerly at xAI.
• Tara Rezaei, founding team member. Formerly an intern at OpenAI.

The senior research and engineering team beyond the named founding cohort has not been publicly disclosed.

Funding and backers

Mirendil's funding history through April 2026 consists of a single round in negotiation but not closed: the reported $175 million at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins. Round terms have not been finalized as of April 2026.

The Andreessen Horowitz lead is consistent with the firm's pattern of leading rounds for high-profile AI Insurgents (alongside its lead in Thinking Machines Lab's $2 billion July 2025 seed). The Kleiner Perkins co-lead is consistent with the firm's track record of investing in scientific-AI ventures including past adjacent rounds.

The company has not disclosed cumulative compute commitments, cloud-partner relationships, or any closed funding rounds. The reported valuation of approximately $1 billion is high relative to typical 2026-vintage Insurgent-lab seed rounds and reflects the founders' scientific-reasoning credentials and the lead investors' belief in the vertical-specialization thesis.

Industry position

Mirendil occupies a distinctive position within the 2025-to-2026 "neo-lab" cohort. The combination of senior Anthropic-and-DeepMind founders, the scientific-domain vertical specialization, and the reported $1 billion seed valuation produces a profile most directly comparable to other vertical-AI labs founded by frontier-lab alumni.

The closest peer comparators are vertical-AI labs that target science as the primary application domain. Periodic Labs, founded by Liam Fedus (former OpenAI VP of Research, Post-Training) and adjacent ex-OpenAI scientific-research leaders, pursues a similar science-AI thesis. Lila Sciences and FutureHouse target adjacent scientific-discovery research surfaces. EvolutionaryScale targets biological foundation models. Isomorphic Labs, the Demis Hassabis-led Alphabet spinout, pursues drug discovery using DeepMind-derived AI capability.

The strategic risks are substantial. The company has not released a model, the technical direction has not been comprehensively disclosed, and the scientific-vertical thesis has not been validated through a shipping product. The valuation depends on the founders' research credentials and the lead investors' belief in the long-term thesis rather than capability evidence.

The strategic strengths are distinctive. Neyshabur's leadership of scientific-AI reasoning programs at both Google and Anthropic provides direct context on what frontier scientific-reasoning capability looks like inside two of the leading frontier labs. The Andreessen Horowitz and Kleiner Perkins lead-investor combination signals strategic-investor confidence at the senior tier. The biology-and-materials-science specialization is differentiated within the broader cohort of general-purpose Insurgents.

Competitive landscape

Mirendil competes with several Frontier, Insurgent, and adjacent vertical-AI labs:

• Anthropic. The founders' previous lab. Anthropic's general-purpose Claude models have demonstrated strong scientific-reasoning capability, and Anthropic has stated public interest in scientific-AI applications. Direct competitor on the scientific-reasoning surface.
• OpenAI and Google DeepMind. Dominant frontier labs whose generalist scientific-reasoning capability is the closest existing comparator to Mirendil's specialized approach.
• Periodic Labs. Closest peer Insurgent on the science-AI thesis. Founded in 2025 by ex-OpenAI scientific-research leaders, pursuing a similar but partially differentiated technical direction.
• Isomorphic Labs. Alphabet-affiliated drug-discovery lab using DeepMind-derived AI capability. The dominant scientific-AI player by capital and pharmaceutical-industry partnership scale.
• EvolutionaryScale, FutureHouse, and Lila Sciences. Specialized vertical-AI labs targeting adjacent scientific-discovery surfaces.
• Pharmaceutical-industry AI labs. Internal AI research programs at Pfizer, Roche, and adjacent pharmaceutical companies. Compete on the same drug-discovery customer base that Mirendil's eventual product surface will likely target.

Outlook

Several open questions affect Mirendil's trajectory in 2026 and 2027:

• The closing of the reported $175 million round, including final lead investors, structure, and post-money valuation.
• The first published research output, including the company's specific scientific-reasoning architecture and any empirical demonstration in biology or materials science.
• The first model release, if any, and the capability profile relative to frontier and vertical comparators.
• The commercial strategy beyond research output, which has not been publicly stated.
• Senior-talent recruitment from Anthropic, Google DeepMind, and the broader academic scientific-AI community.
• Strategic-partner relationships, particularly with pharmaceutical companies and materials-science research consortia.

Sources

• Tech Startups: Ex-Anthropic researchers launch Mirendil, target $175M at $1B valuation for AI-powered scientific discovery. Primary source on the founding details and reported funding round.
• The Decoder: Ex-Anthropic researchers launch AI startup Mirendil to tackle scientific research. Coverage of the founding team and mission.
• MLQ: Ex-Anthropic team unveils Mirendil AI to advance biology and materials science research. Founding coverage.
• Behnam Neyshabur personal site. Career history including TTI-Chicago, NYU, Princeton, Google, and Anthropic.
• Behnam Neyshabur on Google Scholar. Research-publication record.
• Dealroom: Mirendil company information. Reference profile.
• opentools.ai: Meet Mirendil. Founding context.