Ex-Anthropic Lab Says 1 Researcher Built a Model Under $10M

Mirendil, an AI lab founded by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, said on 21 October 2026 that its first pretrained model, Mirendil-Base, “was built by one researcher in 3.5 months, with less than $10M in compute.” The company says that figure covers “all the research and small-scale experiments it took to get there, as well as the final training run,” and that “AI agents wrote over 99% of the code.”

The claims are the company’s own. Neyshabur wrote on X the same day that Mirendil will “share independent evaluations, post-training benchmarks and API-access sooner than you think,” so no outside test of the model has been published yet.

Business Pill · PRETRAINING

A short Business Engineer explainer on pretraining: how an AI model learns before anyone teaches it. It teaches the general idea only and says nothing about Mirendil or its model.

What Mirendil Says It Built

Pretraining is the first and most expensive stage of making an AI model: the model reads a very large body of text and code and learns to predict what comes next. Mirendil’s post says that for most of the past decade this work “has taken a dedicated team: dozens of specialists at open-model efforts, and hundreds at frontier labs.” Its experiment asked whether “one researcher, working with agents” could “pretrain a competitive coding model.”

• Division of labour. “AI agents wrote over 99% of the code, ran experiments and diagnosed failures, while the researcher decided what to try and what to trust, and intervened when needed,” the post says. It does not name the researcher.
• Cost scope. “That figure is all-in: it includes the agents’ token spend and data processing, and there was no distillation and we did not purchase or license any additional data.”
• From scratch, public data. The project “started from an empty repository”, built on “generic libraries like PyTorch, JAX and a handful of others”, and “All of the training data came from GitHub and other publicly available code, web pages and public datasets.”
• Model sizes. “Mirendil-Base is in the 100B–500B parameter range, Mirendil-Base-Large in the 500B–2T range, and Mirendil-Base-XL in the 2T–10T range.” The post says Large is “about 80% of the way through training” and XL “about 19% of the way through.”

What Was Posted on X

Mirendil chief executive Behnam Neyshabur announced the result on X on 21 October. His post says Mirendil-Base was built by one researcher in 3.5 months with less than $10M in compute, gives the model’s size as the 100B-500B parameter range, and says Mirendil will share independent evaluations, post-training benchmarks and API access later. The attached chart is Mirendil’s own comparison on held-out code from its repositories; the piece treats all of these figures as the company’s claims.

How It Compares, by Mirendil’s Own Test

Mirendil compares its model with five coding models on one measure, bits per byte, which the post describes as “the model’s loss per byte of text, so models with different tokenizers can be compared fairly.” It is a measure of how well a model predicts code, not of how well it completes tasks. On seven outside repositories, Mirendil-Base scores better than three of the four comparison models and behind Kimi-K3, according to the post’s Figure 2.

Average row of Figure 2 in Mirendil’s 21 October 2026 post: cross-entropy in bits per byte on seven external repositories published after the models were trained. Lower is better. The test was designed and run by Mirendil; Inkling is in its Figure 1 but not in Figure 2.

The post notes that “Kimi-K3 and GLM-5.3 are post-trained models, which generally score slightly worse on bits per byte than their base versions.” Its Figure 1, scored on “held-out code from Mirendil repositories”, puts the two still-training models, Large and XL, ahead of all five comparison models.

Who Is Behind Mirendil

Mirendil’s mission, in Neyshabur’s post, is “to solve Recursive Self-Improvement (RSI) and use it to democratize frontier AI development.” Recursive self-improvement means AI systems that help build better AI systems, each generation making the next one faster to produce. Mirendil’s blog post says Mirendil-Base “didn’t build itself autonomously” and that “Our next model will be built partly by Mirendil models.”

• Founders. Neyshabur and Mehta “met at Google in 2019, moved to Anthropic in late 2024, then left in December 2025,” The Next Web reported in June.
• Team and seed. Mirendil’s 24 June announcement says its founding team “consists of 20 researchers and engineers from frontier institutions including Anthropic, xAI, Google DeepMind, and OpenAI” and that Andreessen Horowitz and Kleiner Perkins “led our seed round of $200M, followed by an investment from NVIDIA among others.”
• Compute partner. On 6 August Mirendil announced “a multi-year partnership with Google Cloud”, using Google Cloud TPUs and Nvidia infrastructure.
• New round talks. Bloomberg reported on 22 September that Mirendil was in talks to raise new funding at a $5 billion valuation, with Kleiner Perkins in talks to lead a round that could bring in up to $1 billion, according to a PYMNTS summary of that report. We have not seen confirmation that the round closed.

Connecting the Coverage

Mirendil is the third lab in two days to describe AI doing part of the work of building AI. Our Recursive piece reported that Richard Socher’s lab released Flint on 21 October and said “Our small team of people and agents built it from scratch in under four months.” Our Anthropic report piece quoted Anthropic’s 2026 Public Benefit Report: “Claude already performs a growing share of our safety research, and we are increasingly relying on today’s models to train future ones.” That piece also carries The Information’s report that OpenAI is weighing limits on AI systems that can improve themselves.

Our Microsoft proof piece covered a project in which AI agents explored roughly one hundred research trajectories while researchers pruned the search, and our cost-to-outcome piece quoted Satya Nadella saying customers will “measure the cost of tokens against the outcomes they produce.” As we read it, Mirendil’s claim puts a price on one such outcome: a pretrained base model, for under $10 million in compute and one researcher’s time, by its own account.

What It Means for Businesses

As we read it, the claim to watch is the shift in the bottleneck. Mirendil’s post says that “When agents do most of the hands-on work, the bottleneck moves to compute, to data, and to the judgment of whoever is directing the agents.”

As we read it, if independent tests support the result, the cost of a custom base model for a specialist company would be set mainly by compute and data rather than by hiring a pretraining team. The post names “a chip-design team, a robotics lab or a hospital training a model on its own data” as examples. That depends on evaluations Mirendil has not yet published.

The Structural Read

As we read it, the post describes a division of labour rather than an autonomous system: agents wrote the code and ran the experiments, and one person decided what to try and what to trust.

As we read it, the evidence so far is Mirendil’s own: a prediction measure on code, chosen and run by the company, with independent evaluations promised but not yet published.

As we read it, the post places the result inside a wider shift, citing internal measurements from Anthropic and OpenAI that agents are taking on more of the work of building the next generation of AI.

Three Implications

FOR COMPANIES THAT WANT THEIR OWN MODEL As we read it, if outside tests support Mirendil’s claim, the price of a custom base model would be set mainly by compute and data; the post’s own examples are “a chip-design team, a robotics lab or a hospital training a model on its own data.”

FOR FRONTIER LABS As we read it, a lead built on the size of a pretraining team is the part this claim puts under pressure; Mirendil’s own figures, not an independent test, are the evidence so far.

FOR INVESTORS As we read it, the result arrives as Mirendil was reported by Bloomberg, via PYMNTS, to be in talks to raise new money at a $5 billion valuation; that round has not been confirmed as closed.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework’s line on the first moat: “The Corpus Moat — built by pretraining, eroding as the public web is exhausted and synthetic data closes gaps.”

As we read it, Mirendil’s claim is a test of that line: a base model pretrained only on public data by one researcher and agents, by the company’s account. The framework was published in July 2026 and does not discuss Mirendil.

What Is Not Established

No independent evaluation of Mirendil-Base had been published when we wrote this; the comparisons are Mirendil’s own, on a prediction measure it chose. The post gives parameter ranges, not exact sizes, and does not name the researcher. The $10 million figure is for compute; the post gives no figure for salaries or other costs. We could not read Bloomberg’s September report directly and rely on PYMNTS’s summary of it.

The Bottom Line

Mirendil says one researcher, with AI agents writing over 99% of the code, pretrained its first model in 3.5 months for under $10 million in compute. On its own seven-repository test the model scores better than three of four comparison models and trails Kimi-K3. Independent evaluations, which Neyshabur says are coming, will decide how much weight the claim carries.