Anthropic’s IPO Bets on Millions of AI Experts

The next generation of Claude is being built with help from the current one. Anthropic says Claude wrote more than 90% of the contributions to its codebase in September 2026. Engineers increasingly direct and review the work.

On October 26, Anthropic announced its filing for a proposed initial public offering, a sale of shares to public investors. Its prospectus sets out a much bigger ambition. Within a few years, it expects widely available systems that rival leading human experts across most professional and scientific fields, undertake work lasting hours to weeks, and operate connected laboratory equipment and robots. Ultimately, millions of collaborators could work in parallel: a “country of geniuses in a data center.”

This is a forecast, not a description of a workforce Anthropic has already deployed. But it poses a question worth taking seriously now: what changes when the scarce resource in research is no longer necessarily the number of capable minds available to work on it?

What Changes When Expertise Can Be Copied?

A brilliant scientist still has one working day. A lab cannot make a thousand copies of her to investigate a thousand promising leads before breakfast.

A sufficiently capable AI system could be run in many instances, each with a different assignment. One could examine a mathematical approach, another check a simulation, and others search for evidence that the favored idea is wrong. The immediate gain would be a wider search, not merely a faster answer to the same question.

Imagine a small research team with a promising result and a long list of things it cannot afford to investigate. Cheap, reliable expert assistance could change which questions get asked at all. It might make neglected comparisons or laborious checks worth doing. That opportunity depends on access and cost, not simply on the model existing inside a leading company's lab.

There is another important distinction. A thousand copies of the same model are not automatically a thousand scientists with complementary judgment. They may repeat the same assumption. Dividing the work, checking disagreements and testing the results would be part of the research system, not administrative details added afterward.

The interesting promise is therefore not an enormous head count. It is a much larger supply of useful, testable intellectual work.

The Researcher Is Also a Toolmaker

AI research adds a feedback effect to that picture. A model can help improve the tools used to build its successor. If the successor then becomes a better research assistant, the next team starts with better tools again.

Anthropic describes Claude assisting with experiments, training data, evaluations and research code. One intended payoff is finding ways to make more effective use of available computing resources. Better methods could make a future training run more productive without an equivalent increase in hardware. The company calls this development cycle “closing the loop.”

Consider a narrower, hypothetical example. A researcher wants to test a new training method, but preparing and debugging the experiment takes most of the week. An agent that handles that work reliably could let the researcher compare more ideas. If one comparison produces a genuine improvement, the next round could benefit. No autonomous scientific mastermind is required for that first gain.

The September coding figure is evidence of extensive use, not a measurement of how much research has been automated. Writing code does not mean choosing most of the worthwhile experiments. In its research account of self-improvement, Anthropic identifies goal selection and judgment as important remaining gaps. Fully autonomous development of successor models remains an uncertain endpoint.

Faster cycles also need trustworthy tests. Otherwise, a system can carry a mistake forward as readily as an improvement. The question of who can inspect and redirect that loop becomes more important as the work accelerates.

A Search Through DNA Shows the Possibility

The most tangible illustration comes from a separate research project, not from the IPO filing.

On September 23, Anthropic reported a biological discovery initiated by Claude agents. Scientists had asked them to search DNA databases for interesting examples of reverse transcriptases, a class of enzymes that copy RNA into DNA.

The agents noticed an unusual arrangement beside a reverse-transcriptase gene: repeating DNA sequences alongside an additional protein-coding gene. The pattern pointed to a previously uncharacterized system in viruses that infect bacteria. The researchers named it ART, short for array-associated reverse transcriptases.

This is an appealing glimpse of the proposed future: machines combing through an enormous body of existing information and finding something humans can investigate. In the preprint, the discovery campaign involved 949 agent sessions and ran for 21.5 hours. That is a count of software sessions, not 949 independent scientists.

Human scientists performed the laboratory work. Experiments showed short RNA products associated with the repeating sequences. The system's biological function remains unknown, and the work did not establish a usable gene-editing tool.

The paper also contains a revealing result: ten further search campaigns failed to rediscover the array. That does not erase the first discovery. It shows why a successful episode and a dependable discovery process are different achievements.

The opportunity is real enough to investigate: search more of biology, notice more unusual patterns, bring better candidates to the lab. The next challenge is turning an exceptional find into a method scientists can rely on.

Weeks of Work Require More Than a Running Agent

The prospectus's long-duration vision is easy to picture: assign a difficult project, come back later, receive a finished result. The difficult part is preserving good judgment through everything that happens in between.

A project can encounter missing information, misleading results or a change in priorities. The system must know when to continue, when to abandon an approach and when to ask a person. A program remaining active for weeks establishes none of those things by itself.

Even the language used to measure progress can mislead. In METR's task-horizon evaluations, duration refers to how long a human expert would take to complete the tasks, at a specified estimated AI success rate. It is not a measurement of how long the AI runs unattended. The tasks are also concentrated in well-defined software and related technical work, not the full range of professional life.

For a lab or a business, the practical question is what proportion of important projects finish correctly, how much intervention they require, and what the whole process costs. A longer runtime is useful only when it carries worthwhile work toward completion.

The Race Moves From Ideas to Experiments

If expert assistance becomes abundant, its output will meet the rest of the world. A proposed material still has to be made. A biological result still has to survive an experiment. Some questions can be investigated in parallel; others depend on what the previous test reveals.

Robotics could help connect the two worlds, but access to a machine is not the same as dependable physical work. Instructions must fit the instrument, the sample and the circumstances. Even Anthropic's ART team did its experimental work with people, rather than an autonomous laboratory.

Computing capacity is another constraint. Running many capable agents requires resources that could also serve customers or train models. Improving efficiency could expand what a lab can attempt, while an expanding appetite for experiments could consume those gains. “More efficient” need not mean “less demand.”

This gives the IPO vision two very different possible outcomes. Reliable, accessible expert assistance could let smaller teams investigate questions that currently require much larger organizations. Or the best systems and the infrastructure needed to exploit them could remain concentrated in a few hands. Capability alone does not decide who benefits.

The milestone worth watching is not the arrival of a million running agents. It is a research team using them to ask better questions, test more of them and produce findings that hold up. That would be a profound change even before any machine could run the entire lab.

Anthropic is asking the public to imagine expertise becoming plentiful. The deeper test is whether dependable discovery can become plentiful with it.

For the workplace version of that question, read why more AI working hours do not automatically mean more finished work.