Signal › Feed

61% of Enterprises Underestimated First-Year AI Costs by 40%. The 2026 Consumption Pricing Shock and the Governance Framework That Prevents It.

Former Anthropic researchers Behnam Neyshabur and Harsh Mehta closed a $1 billion Series A on November 12 to scale Lor, an agent platform that coordinates fleets of AI for scientific and engineering research.


On November 12, 2026, Mirendil closed a $1 billion Series A led by Kleiner Perkins and Clearlake Capital. The round included participation from Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, and S32. No post-money valuation was disclosed; sources cited by multiple outlets suggested approximately $5 billion including the new capital.

This came five months after a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins, with Nvidia also participating. Total capital raised: $1.2 billion. Speed from seed to $1 billion Series A: approximately 140 days.

That velocity is not incidental. It is the investment community's answer to a specific question: who builds the tools that build the next generation of AI? Mirendil's bet is that the answer is not more human researchers alone — it is fleets of AI agents coordinating under human direction, running experiments and iterating on research at a pace no team of humans can match. The $1.2 billion is the market's first large signal that the bet is credible.

The Founders: Anthropic Veterans With a Specific Research Pedigree

Mirendil was co-founded by Behnam Neyshabur (CEO) and Harsh Mehta (CTO), both former Anthropic researchers who departed in December 2025. The founders' backgrounds are directly relevant to what the company is building.

Neyshabur's background spans two of the leading research institutions in AI. Reports describe him variously as having co-led Anthropic's Discovery team — the internal team responsible for Anthropic's most significant research breakthroughs — and Google DeepMind's Blueshift team, which worked on automated research capabilities. The specific institutional home is disputed across sources, but the substance is consistent: Neyshabur worked at the frontier of using AI systems to accelerate AI research itself.

Mehta's background is specifically relevant to Mirendil's commercial product. He is credited with building the first version of Anthropic's internal research automation platform — the tooling Anthropic used to run experiments faster and more autonomously than traditional research workflows allowed. Mirendil is, in a meaningful sense, Mehta's attempt to commercialize for the broader market what he built internally at Anthropic.

The a16z investment thesis for the seed round framed this as a bet on specific insight: the people most capable of building AI-led research tools are those who have already built them inside leading labs, understand the failure modes from production use, and know what the gap between internal tooling and a commercially scalable product actually looks like. That framing remains the core credibility claim for a company that has not yet published benchmarks, customers, or revenue.

The Round Structure: Why $1 Billion at This Stage

Series A rounds are historically $5–50 million bets on early traction. Mirendil's round is sized like a Series D for a company with no announced revenue. Understanding why requires understanding what the company needs the capital for.

The frontier AI development market has a specific capital requirement that does not scale with company maturity: compute. Training models in the 100 billion to 500 billion parameter range — where Mirendil's proprietary Mirendil-Base model is reported to sit — requires compute commitments that dwarf most early-stage company capital needs. The multi-year Google Cloud deal, reported at over $100 million and providing access to TPUs and NVIDIA GPUs, illustrates the infrastructure cost of operating in this category.

The growth equity logic is therefore: a first-mover on agent-led AI research infrastructure needs to lock in compute infrastructure before competitors do. Capital requirements at this tier are not proportional to company maturity — they are proportional to the category. Kleiner Perkins and Clearlake are not betting on Mirendil's current revenue. They are betting on the structural advantage that large-scale compute access creates for a company whose tools themselves are AI research systems.

The investor composition adds a signal layer. Google's direct participation — alongside its separate AI research organization in DeepMind and its own frontier model development — is not a commodity check. When Google invests in a competitor's research automation platform, the interpretation is either strategic hedge or genuine conviction that the platform addresses capabilities Google doesn't want to build internally. The more likely read: research automation infrastructure serves a market that benefits Google's broader AI ecosystem, and the value of the tool compounds with scale across the organizations using it.

Lor: The Agent Platform in Private Beta

Mirendil's commercial product is Lor — an agent platform in invite-only private beta as of the November Series A.

Coverage of the Lor platform describes it as an agent orchestration system where individual agents write code, run computational experiments, analyze results, and iterate on research directions over multi-day sessions without continuous human supervision. Users define a research objective; Lor decomposes it into discrete experiments, assigns each to agents, monitors execution, and surfaces findings to human researchers who set the next direction.

The model architecture is flexible by design. Lor runs on Anthropic's and OpenAI's frontier models alongside open-source models and Mirendil's own proprietary models. Different research sub-tasks have different requirements: a code-writing sub-task might run best on a model optimized for coding; a long-context analysis sub-task might use Fern, Mirendil's own long-context model; a hypothesis generation step might draw on the strongest available frontier model. Locking the platform to a single model vendor would limit the task range.

The absence of pricing and general availability is consistent with the early-stage private beta approach: Mirendil needs to understand the structure of enterprise research workflows — data access patterns, verification requirements, security constraints, compute costs per experiment type — before it can design a pricing model that captures value appropriately. The invite-only beta is as much a data-gathering exercise as a product validation exercise.

The Central Claim: One Researcher, Hundreds of Agents, Less Than $10 Million in Compute

The founding data point at the center of Mirendil's Series A narrative is specific and extraordinary: one Mirendil researcher, working with a fleet of AI agents and less than $10 million in compute, built Mirendil's first foundation model.

If this claim holds up at scale, it represents one of the most significant shifts in the economics of AI development since the transformer architecture itself. The conventional model for training frontier AI systems involves large teams, massive compute budgets, and years of cumulative research progress. DeepSeek's demonstration that strong capabilities could be achieved at substantially lower compute cost than previously assumed disrupted the hardware moat narrative. Mirendil's claim goes a step further: not just cheaper compute, but fewer humans, substituting agent labor for researcher time.

The implications, if the claim generalizes:

DimensionTraditional Lab ModelMirendil's Agent-Led Model
Human researchers per modelDozens to hundredsSmall team + agent fleet
Compute for first foundation model$100M+ historicallyUnder $10M (claimed)
Experiment cycle timeWeeks per major iterationDays (continuous autonomous loops)
Human bottleneckResearcher attention and calendarCompute budget and agent orchestration
Research agenda scopeLimited by researcher bandwidthParallelizable across agent fleet
Verification requirementPeer review and replicationStill requires human review

It is worth being precise about the epistemic status of this claim. Mirendil's $10 million figure for its first foundation model is the company's own account, reported without independent verification. No technical paper has been published. No external benchmarks have been reported. The claim exists in the public record as investor narrative, not peer-reviewed science.

The investment signal is not nothing. When Google, Kleiner Perkins, Clearlake Capital, and Nvidia collectively put $1.2 billion behind a team and a thesis, those investors have done the diligence the public has not. The absence of public verification does not mean the claim is false. It means the claim is unverified — the normal state for a company in private beta.

The Competitive Landscape: Where Mirendil Sits

The obvious comparison is to AI coding agents — Cognition AI's Devin, OpenAI's o3 and o4 deep research capabilities, and the emerging category of autonomous software engineering agents. That comparison understates the scope.

Coding agents automate execution of well-defined software tasks: implementing specified features, writing tests, refactoring code. The input is a human specification; the output is code that satisfies it. Research automation addresses a different problem: generating hypotheses, designing experiments, interpreting results, and updating the research agenda. This requires scientific judgment at every stage — which hypotheses are worth testing, which experimental designs are valid, which findings are meaningful versus artifacts of the measurement approach.

Signal's earlier coverage of Thinking Machines Lab's $5 billion raise documented the emerging competitive landscape among well-capitalized labs pursuing different infrastructure theses for the next generation of frontier AI development. Mirendil's distinctive claim is that competitive advantage in model development will increasingly come not from raw compute scale alone but from research throughput — running more experiments faster — and that agent-led research automation is the primary lever for increasing that throughput.

This positions Mirendil against an unusual competitive set: not consumer AI products, not enterprise workflow automation, but the internal tooling of the world's leading AI labs. The company is claiming it can commercialize a capability that most frontier labs have been building privately, and sell it back to those labs and to the broader R&D enterprise market.

What Agent-Led R&D Means for Enterprise Research Organizations

Mirendil's immediate market is AI research teams. But the thesis generalizes to any organization that runs structured research processes — pharmaceutical companies, semiconductor designers, materials science labs, and software product teams building increasingly complex AI-powered products.

Drug discovery maps closely to what Lor describes. Identifying candidate compounds, synthesizing them, testing in cell lines and animal models, analyzing failure modes, and updating the hypothesis space based on results involves the same decompose-assign-analyze-iterate loop that agent-led research automation addresses. The constraint in pharmaceutical R&D is not that the experiments are hard to reason about — it is that they take months and cost millions at the wet lab stage. Agent automation reduces the human coordination overhead but cannot eliminate the physical experimental cost. The leverage is highest on the computational stages: virtual screening, protein structure prediction, activity-off-target prediction, formulation modeling.

The Anthropic Claude Duet proactive agent model demonstrated how autonomous agents can run sustained multi-day workflows without continuous human supervision — the same capability Lor requires. Research automation that runs overnight without human intervention, surfacing results for researcher review in the morning, is a qualitatively different workflow from research automation that requires human supervision at each step.

For enterprise R&D leaders evaluating agent-led research automation, four operational questions matter:

1. Experiment decomposability. Does your research workflow decompose into discrete sub-tasks with measurable outcomes? Agent-led automation works best when the research process can be broken into well-defined experiments. Highly exploratory research that requires deep tacit judgment at every step is a weaker fit.

2. Verification infrastructure. Every finding generated by autonomous agents needs human review before it enters scientific or commercial decision-making. Organizations without the infrastructure to review AI-generated research at scale will create new failure modes, not reduce them.

3. Compute cost modeling. For computational experiments, agent automation has strong unit economics — agents run continuously at low marginal cost per experiment. For wet lab experiments, the agent reduces coordination overhead but not experimental costs. The economic case depends on the mix.

4. Data governance. Research automation systems have access to your proprietary research data, hypothesis space, and experimental infrastructure. The security and IP ownership model for Lor must be evaluated carefully for organizations in regulated industries or with significant patent-based competitive advantages.

The Trust and Verification Problem

The deepest structural challenge in agent-led AI research is not capability — it is verification. Autonomous AI systems doing scientific research can produce results that look valid but contain subtle errors: training artifacts, benchmark overfitting, spurious correlations in experimental data, emergent reward-hacking in reinforcement learning setups.

Human research organizations have institutional processes for catching these problems: peer review, replication requirements, independent evaluation across teams. These processes are slow, but they exist because they catch real failure modes that internal teams miss.

Research automation systems need equivalent processes built into their architecture. For a system like Lor, this means: how does the platform distinguish a genuine experimental finding from an agent that optimized toward the appearance of a finding? How does it handle conflicting results across agents running parallel experiments? What is the provenance model for a finding that was generated by 50 agents across 500 sub-tasks — and what can a human researcher verify independently?

None of these questions have published answers for Lor specifically. They are the right questions to ask before deploying agent-led research automation for commercially significant work, and they represent the gap between what the platform claims and what enterprises will require before trusting its outputs.

The enterprise AI governance infrastructure debate of 2026 — documented in LangChain's $300 million Series C — established that production AI agents require governance infrastructure at the execution layer. Research automation adds a harder requirement: governance at the scientific validity layer, which is a different and more difficult problem than governing whether an agent accessed the right data sources.

What Investors Are Actually Betting On

The Series A investor composition includes Google directly alongside Kleiner Perkins and Clearlake Capital. Google's direct investment is worth noting specifically: Google has its own world-class AI research organization in DeepMind, its own frontier model training infrastructure, and its own agent development programs. When Google invests in a company building research automation tools, the signal is either strategic hedge or genuine conviction that the platform addresses capabilities Google hasn't built or doesn't want to build internally.

The more likely interpretation: research automation infrastructure is not a winner-take-all market at the frontier, and the value of the tool compounds with scale — the more research organizations using Lor, the more data about research workflows that can improve Lor's research execution capabilities. Google is investing in a platform that will be used by organizations that are not Google, in research workflows that benefit Google's broader ecosystem.

Kleiner Perkins and Clearlake's decision to co-lead at $1 billion is a growth equity signal. Kleiner Perkins has historically made growth equity bets on companies with demonstrable traction — the Series A scale implies a view that Mirendil's private beta results justify treating this as a growth equity investment rather than an early-stage venture bet. Clearlake Capital's industrial and technology focus adds a different dimension: Clearlake looks for operational improvements at scale, which is consistent with a thesis that research automation can be deployed across large enterprise R&D organizations.

The Road to General Availability

Mirendil has not published a GA timeline for Lor. The invite-only private beta approach suggests the company is prioritizing depth over breadth in its initial customer relationships — understanding exactly how research organizations use the platform, where it fails, and what verification and security requirements enterprise R&D environments impose.

The Google Cloud partnership provides the compute infrastructure needed to run the beta at scale. The $1 billion in capital provides the runway to iterate on the product without revenue pressure while the beta generates the usage data needed to design the pricing model.

The expected path to GA involves several known gates: publishing model benchmarks to give the market a performance baseline to evaluate, expanding the beta to cover more research workflow types, establishing the IP ownership and data governance model for enterprise deployment, and building the verification tooling that lets human researchers audit agent-generated findings efficiently.

Takeaway: Mirendil's $1 billion Series A is the largest bet yet on the thesis that AI agents can do AI research itself — not just execute well-defined tasks, but run the hypothesis-experiment-analysis loop that has traditionally required teams of PhD researchers. The founding team's direct provenance from Anthropic's internal research automation work gives the bet more credibility than a typical moonshot raise. What remains unverified — third-party benchmarks, published model capabilities, production customers — will matter as much as the capital for the long-term outcome. For enterprise R&D leaders, the implication is clear: watch Lor's progress toward general availability closely, and start building the internal verification infrastructure that agent-generated research findings will require before they can safely enter commercial decision-making.

Frequently Asked Questions

What did Mirendil raise in its Series A and who invested?

Mirendil closed a $1 billion Series A on November 12, 2026, led by Kleiner Perkins and Clearlake Capital. Additional investors in the round include Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, and S32. No post-money valuation was disclosed with the announcement, though reporting from multiple outlets cited a figure of approximately $5 billion including the new capital. The Series A followed a $200 million seed round announced on June 24, 2026, led by Andreessen Horowitz and Kleiner Perkins, with Nvidia also investing in that earlier round. Total capital raised since founding reaches approximately $1.2 billion. The speed from seed round to $1 billion Series A — roughly 140 days — is among the fastest progressions from seed to that capital scale in the AI sector's recent history.

Who are Mirendil's founders and what is their background?

Mirendil was co-founded by Behnam Neyshabur (CEO) and Harsh Mehta (CTO), both former Anthropic researchers who departed in December 2025. Neyshabur previously co-led either Anthropic's Discovery team or Google DeepMind's Blueshift team — sources disagree on the specific role, but both are described as research automation initiatives at frontier AI labs. Harsh Mehta's background is specifically relevant to Mirendil's commercial product: he is credited with building the first version of Anthropic's internal research automation platform, the tooling Anthropic used to run experiments faster and more autonomously. Mirendil is Mehta's commercial version of what he built internally, applied to organizations outside of Anthropic. The founders' direct provenance from internal research automation work at leading AI labs is the central credibility claim underpinning the company's outsized seed and Series A raises.

What is Lor and what does it do?

Lor is Mirendil's agent platform, currently in invite-only private beta. It coordinates fleets of AI agents that execute research and development tasks autonomously: writing code, running computational experiments, analyzing results, updating hypotheses, and iterating on research directions over multi-day sessions. Users define a research objective; Lor decomposes it into discrete sub-tasks, assigns those tasks to specialized agents, monitors agent progress, and surfaces findings to human researchers for review and direction-setting. Lor runs on multiple model layers simultaneously — Anthropic's and OpenAI's frontier models, open-source models, and Mirendil's own proprietary models including Fern (a long-context model) and Mirendil-Base. The multi-model architecture is deliberate: different research sub-tasks have different requirements, and locking research automation to a single model vendor limits the range of tasks the platform can handle effectively. No pricing has been published; the invite-only beta allows Mirendil to understand enterprise research infrastructure requirements before designing a commercial model.

What is the significance of Mirendil's claim that one researcher built its first foundation model for under $10 million?

Mirendil has stated that one researcher, working with a fleet of AI agents and less than $10 million in compute, built Mirendil's first foundation model. If verified, this would represent a significant data point in the economics of AI development. The frontier model development orthodoxy holds that training frontier-capable models requires hundreds of researchers, thousands of GPUs, and hundreds of millions in compute. DeepSeek R1 disrupted the compute assumption by achieving strong capabilities at a fraction of previously assumed cost. Mirendil's claim goes further — fewer humans, not just less compute — by asserting that agent-led research automation can substitute for a substantial portion of the human researcher bandwidth traditionally required. The claim is currently unverified: Mirendil has not published a technical paper, independent benchmarks, or third-party validation. The investor conviction — including from Google and Nvidia, who have deep AI development expertise — represents indirect validation, but the public record rests on the company's own account.

What enterprise organizations should evaluate Mirendil's Lor platform?

Organizations most likely to benefit from Lor fall into three categories. First, AI research teams — companies building foundation models or running large-scale ML research who need to run more experiments per unit of researcher time. Second, pharmaceutical and biotech R&D teams — drug discovery and compound optimization involve systematic exploration of large hypothesis spaces that map closely to what agent-led research automation addresses, though wet lab experiments add external costs that agent orchestration cannot reduce. Third, semiconductor and materials science organizations — computational design problems with well-defined evaluation metrics are strong candidates for autonomous agent iteration. The common prerequisite across all categories is the ability to verify and validate AI-generated research findings before they enter commercial or scientific decision-making. Organizations that lack verification infrastructure will create new failure modes by deploying research automation, not reduce them.

How does Mirendil compare to AI coding agents like Devin or OpenAI's deep research?

Mirendil's Lor platform addresses a broader problem than AI coding agents. Coding agents like Cognition's Devin automate execution of well-defined software engineering tasks: implementing specified features, writing tests, refactoring existing code. The input is a human-defined specification; the output is code that fulfills it. Research automation addresses the upstream problem: generating hypotheses, designing experiments to test them, interpreting results, and updating the research agenda. This requires not just competent execution but scientific judgment — which hypotheses are worth testing, which experimental designs are valid, which findings are meaningful versus statistical artifacts. OpenAI's o3 and o4 deep research capabilities address research synthesis and analysis in well-documented domains; Mirendil's claim is to orchestrate original experimental research workflows that generate new data, not synthesize existing knowledge. The scope of what Mirendil claims to automate is significantly larger than what current coding agents or research synthesis tools address.