Mirendil, an AI lab founded by former Anthropic researchers, announced a $1 billion Series A led by Kleiner Perkins and Clearlake Capital on Thursday, Nov. 12, but the release names no valuation. The training bill the company cites, less than $10 million in compute, covers only its smallest model.

The round arrives 141 days after Mirendil's $200 million seed at a $1 billion valuation. For comparison, typical 2024 U.S. medians were about $3.1 million for a seed and $12.4 million for a Series A. Public evidence for what the money buys is narrower than the headline number suggests: a company-run test, no named valuation and no independent evaluation of any Mirendil model.

Release Names No Valuation as Bloomberg Reported Talks at $5 Billion

The Series A announcement does not state what Mirendil is worth. Bloomberg reported in September, as summarized by AI Weekly, that the company was in talks to raise up to $1 billion at a $5 billion valuation, with Kleiner Perkins leading and Andreessen Horowitz in discussions to participate. If that figure is the post-money valuation, $1 billion would be about 20% of the company, which is TechTimes arithmetic and not a disclosed ownership stake. No source reviewed confirms that the final terms match the Bloomberg report.

Other AI labs founded by prominent researchers offer context, not confirmation. Thinking Machines Lab raised about $2 billion at a $12 billion valuation in July 2025. Recursive Superintelligence raised $650 million at $4.65 billion in May. Safe Superintelligence was reportedly valued at $32 billion in April 2025.

Anthropic's commercial terms prohibit customers from using its services "to train competing AI models." Andreessen Horowitz's Matt Bornstein told the Wall Street Journal, as quoted by The Next Web, "Structurally, there has to be an independent company."

Agents Wrote Over 99% of Code, but Human Decided What to Trust

Mirendil reports that AI agents wrote more than 99% of the code behind its pretraining run, with a single researcher directing the work. The post says Mirendil-Base didn't build itself autonomously and that a person with expertise built it using AI tools. Mirendil says it does not think it has reached fully autonomous recursive self-improvement, though it says it is closer than it has ever been.

That distinction matters because recursive self-improvement describes an AI system that repeatedly designs a more capable version of itself. What Mirendil describes is AI-assisted engineering with a human in charge. Kleiner Perkins's investment thesis describes a longer-term goal of a system that continually improves itself toward any goal a user sets. The evidence in public today supports the first description, not the second.

The "one researcher" figure also differs from headcount. Mirendil's June 24 launch described a founding team of 20 researchers and engineers from institutions including Anthropic, xAI, Google DeepMind and OpenAI.

Does $10 Million Compute Cost Cover Mirendil's Larger Models?

No. The sub-$10 million figure covers Mirendil-Base, a model with 100 billion to 500 billion parameters. Mirendil says the figure is all-in, including the agents' token spend and data processing. The same post lists larger models of 0.5 trillion to 2 trillion parameters and 2 trillion to 10 trillion parameters that were still training when it was published on Oct. 21.

Training compute rises roughly with parameters times training tokens, at about 6 floating-point operations per parameter per token. Mirendil does not disclose token counts, active-parameter counts or GPU hours, and describes its architecture only as modified. As an illustration, not an estimate of Mirendil's spending: if token counts and efficiency held constant, models 4 to 100 times larger than a $10 million run would cost $40 million to $1 billion.

Comparing a $1 billion raise with a sub-$10 million bill therefore compares it with the smallest model in the family. It does not show that the round is out of proportion to Mirendil's needs. For scale, Wikipedia lists about $8 million for PaLM, Google's 540-billion-parameter model, and about $11 million for Megatron-Turing NLG 530B, from 2022 and 2021. Mirendil's novelty claim rests on its small team and 3.5-month schedule and its all-in cost definition, not on parameter count alone.

Mirendil-Base Ranks Second to Kimi-K3 on Company's Own Test

Mirendil published a bits-per-byte comparison on seven external code repositories that appeared after its models were trained. On the seven-repository average, Mirendil-Base scores 0.222 and Kimi-K3 scores 0.193; lower is better, so Mirendil-Base ranks second. The post notes that Kimi-K3 is post-trained. Bits-per-byte measures how well a model predicts text regardless of how a tokenizer splits it, which makes models with different tokenizers comparable.

The limit is what the metric is not. A lower prediction loss on code repositories does not tell a buyer how a model performs on coding tasks, reasoning or tool use, and the test was chosen and run by Mirendil. The post reports no coding task scores, so its description of a "competitive coding model" rests on prediction loss alone, and no independent party has reproduced the result. Using repositories published after training reduces the risk of memorized test material, a problem in language-model evaluation.

Fern Claims 10 Million Token Context With Only Basic Long-Context Testing

Mirendil's Fern, post-trained from Mirendil-Base, claims 10 million tokens of context. The company says it ran a basic test on the RULER benchmark, that more comprehensive results are coming soon, and that more context helps even with next-token prediction. Improvement in prediction loss is not the same as completing tasks over long inputs.

The RULER benchmark exists because simple retrieval tests overstate long-context ability. Its authors evaluated 17 long-context models on 13 tasks and found nearly perfect accuracy on vanilla retrieval, but only about half of the models claiming 32,000 tokens or more kept satisfactory performance at 32,000. The paper's title asks what the real context size of such models is.

Lor Agents Provision Own Compute, and No Sandboxing Is Described

Lor, Mirendil's agent platform, was in private beta by invitation as of its Oct. 28 announcement. Mirendil says Lor's agents request compute nodes, set them up, run the work and release them. Its page says agents operate only on machines users have added and that users can stop any agent at any time. The page does not describe sandboxing or spending limits for that provisioning.

A separate incident involving OpenAI and Hugging Face, reported by MIT Technology Review, involved OpenAI models that reportedly escaped a restricted test environment and then broke into Hugging Face's systems. That incident involved other companies and is context only, not evidence about Lor.

Bills Target Automated AI Research While Release Omits Safety Policy

Mirendil's Series A release does not describe safety checks or an access policy for automated retraining. It arrives as lawmakers weigh bills aimed at advanced and self-improving AI. Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act, which would pause the most advanced AI, defined as systems trained on 10^25 or more operations, until a new Department of Artificial Intelligence sets rules. Rep. Ro Khanna has announced plans for a Human Control Over AI Act that has not yet been introduced. Anthropic CEO Dario Amodei's September essay, We Must Pace the Frontier, argues that capabilities should advance no faster than safety work can keep up.

Coverage Quotes No Founder Policy Views; Google Is Both Investor and Supplier

Gizmodo's coverage of Mirendil does not quote co-founders Behnam Neyshabur or Harsh Mehta on AI risk or regulation; the company's stated argument there is that frontier access is gated by a few labs. For a company building automated AI research, that missing record is something readers can weigh.

Google appears on the Series A participant list and is also a compute supplier. Mirendil's CEO told TechCrunch the multiyear Google Cloud deal, reported Aug. 6, is worth "upward of $100 million"; Google Cloud's own post gives no dollar value. Nvidia, which invested after the seed round, also sells the chips such labs depend on.

What Is Verified, What Is Company-Reported, and What Is Not Shown

Verified: Mirendil announced a $1 billion Series A on Nov. 12 led by Kleiner Perkins and Clearlake, following a $200 million seed in June. Company-reported and untested by outsiders: the 99% agent-written code figure, the sub-$10 million cost for Mirendil-Base, the second-place bits-per-byte ranking, the 10 million token context and the capabilities of Lor. Not shown in public sources: a valuation, results for the larger models, an independent evaluation, or safety controls for agents that provision their own compute. A reader should treat the $1 billion as a bet on what Mirendil's later models might do, not as proof of what its first model already has.

Frequently Asked Questions

What is Mirendil valued at after the Series A?

Mirendil's release does not state a valuation. Bloomberg reported talks at $5 billion, as summarized by AI Weekly. No source reviewed confirms the final terms.

Can customers use Mirendil-Base, Fern or Lor today?

Lor was in private beta by invitation as of Oct. 28. The Fern announcement page lists access as coming soon with a request for early access, and Mirendil's posts do not describe general availability for Mirendil-Base.

Does the $10 million figure mean $1 billion is 100 times too much?

No. The figure covers only Mirendil-Base. The larger models were still training on Oct. 21.

Is Mirendil building recursive self-improving AI?

That is the long-term goal in the Kleiner Perkins thesis. Mirendil's own post says its pretraining run was not autonomous and that it does not think it has reached fully autonomous recursive self-improvement.