Anthropic's AI writes 80 percent of its code and leads 26 percent of its research

Anthropic, one of the companies building the most capable AI systems, has published a page about how much of its own work its AI now does. Its title is When AI builds itself, and its most quoted figure is this: as of May 2026, “more than 80% of the code we merge into Anthropic’s codebase” was written by the company’s own AI model. Before February 2025, the page says, the figure was “in the low single digits.”

That is a large change in a short time, and it is easy to read the number as an answer to a bigger question: how much of the work of building AI is now done by AI? It is not an answer to that question. It measures one part of the work, the part that is easiest to hand over, and the same company has published two other numbers that measure more of it.

Three Numbers From One Company

The When AI builds itself page says exactly what the 80 percent figure measures: the share of lines of code that end up in the company’s working software and can be traced to its AI. The company’s leaders have put the share higher, at “90% or more,” counting scripts and experiments as well.

The same company has published two other numbers about the same work.

The first comes from what Anthropic calls a prototype index of how much of its AI research its own AI does. The company sampled about 15,000 tasks from its staff’s working weeks in July 2026, sorted them into a few hundred kinds of work, and graded each kind by how much the AI did, with the kinds that took more staff time counting for more. For August 2026 the result is that the AI “leads” 26 percent of the work, up from under 1 percent in February 2026. “Leads” has a definition: the AI “can complete most of the task end-to-end from a high-level prompt, while the human supervises.” The grading was done by the company’s own AI, which the index’s own page names as a weakness: the judge “could make the same kinds of errors” as the system being judged. The company checked the AI’s grades against its own staff’s, and reports that the two agreed about as often as the staff agreed with each other.

The second is an internal poll published on October 16th, 2026. In August, 1,334 Anthropic staff were asked what share of the work of building an early-2026 frontier model is done by AI today. The median answer was 51 percent. The poll’s page is careful to call these “personal estimates,” and it notes that this is “a broader question than how much code AI writes.”

So one company, in one year, gives 80 percent of the code, 26 percent of the research work led by the AI, and 51 percent of the work in its staff’s opinion. None of these numbers contradicts the others, because they measure different things.

What My Tracker Scores

I keep a public website, the AGI Zeitgeist Tracker, that asks one question: are we entering what people call the singularity, the moment when machines get smarter so quickly that the world changes faster than anyone can keep up with? It answers with four tests, all of which would have to be met at once before it would say that we have entered the singularity. One of them is whether AI is doing most of the work of improving AI, which the tracker sets at half of that work. The tracker scores that test with the 26 percent figure, which puts it 52 percent of the way to its line of one half. On the tracker’s front page, the overall answer, as of October 21st, 2026, is “Not yet.”

If the tracker used the 80 percent figure instead, that test would already be passed, since 80 is above one half. It does not, because the test asks about the work of improving AI, and the code is only part of that work. The tracker lists the 80 percent figure in its table of what the companies have said, where a reader can see it, but it does not count toward the answer. Writing the code is one part of building an AI system. Someone also has to decide what to build, design the experiment, decide whether the result can be trusted, and decide what to try next. The company’s own poll page says as much in a footnote: building a model “also involves research direction, experiment design and evaluation, and more.” A company can reach a point where the machine types most of the lines while people still do most of that deciding. The 26 percent figure tries to count the whole task. The 80 percent figure counts only the part that is easiest to hand over.

What the People Do Now

The When AI builds itself page is plain about what the people at the company now do. In engineering, it says, “humans supply the goal, but they no longer need to supply the method.” In one study it describes, AI agents did far better than two human researchers at a research problem, but “humans still chose the problem and created the scoring rubric,” the rules for judging what counted as success. Another passage expects people to “stop writing code entirely, and shift to only reviewing it,” and warns that if people cannot review code as quickly as the AI writes it, “human review will become the bottleneck.” The company now runs an automated reviewer over changes before they merge, and the same page says that reviewer, looking back, “would have caught roughly a third of the bugs behind past incidents.”

I argued in Specification Is Governance that once a machine can do the work for almost nothing, the work itself stops being where the power is. The power moves to whoever writes the instructions the machine has to satisfy: the checklist, the test, the definition of a job done. The When AI builds itself page shows the same move inside the company that builds the models. The people there set the goal, write the scoring rule and check the output. The 80 percent figure leaves those jobs out.

A second tracker of mine, The Belt Equation, had to settle the same question about a different kind of work. One of its steps is “A non-human mind produces knowledge beyond us”, and I had to decide what would count as the machine’s own idea. The answer I came to is this: “when the operators built the search frame and the scoring rule, then the idea is theirs, even if the system found the answer.” By that rule, several machine-found results in mathematics that beat the best human ones do not yet count as the machine’s ideas. The same rule, applied to code, says that the 80 percent figure tells us only whose lines they were. Whether the work was the machine’s is a different question, and it turns on who set the problem and who judged the answer.

What Would Change the Picture

None of this makes the 80 percent figure small. A company whose engineers no longer write most of their own code has changed how it works, and Anthropic’s When AI builds itself page expects the share to rise. The question for the tracker is a narrower one: when does the machine do most of the deciding, and not only most of the typing?

Two things are worth watching. The first is the 26 percent figure itself. If the company keeps publishing it and it passes one half, that part of the tracker’s test is met on the company’s own measure, and on current trends the tracker expects the figure to cross one half in 2027. The index is graded by the company’s own AI, so I would also want a check from outside the company before relying on it. The second is review. If checking the work also passes to machines, as the automated reviewer suggests it is starting to, then the part left to people narrows to choosing the goal and deciding what counts as success. The When AI builds itself page does not expect that part to stay with people for long. In a section headed “What if we’re wrong?” it says that most progress in AI is routine work between rare big ideas, and that routine work is what the AI now does. It also reports a test on moments where one of the company’s researchers had taken a wrong turn: shown the work up to that point, the AI’s suggested next step was judged better than the person’s 64 percent of the time as of April 2026, up from 51 percent five months earlier. That is one company’s test, judged by its own AI, on moments picked because the person had erred, and the When AI builds itself page says all of that itself. The same page says that judgment, which it calls research taste, is still “an area of human comparative advantage, for now,” and that it “might be just another AI capability that AI systems fail at for a time, then get good at.”

Reading List & Conceptual Lineage

From Sentient Horizons

Specification Is Governance
The argument that as machines take over the doing, power moves to the people who write the rules the machines must meet. This piece finds the same move inside an AI company.

A Rule Against Cruelty, With No Reason Given
The previous Frontier Dispatch, on another of the same company’s decisions about its models.

The trackers

The AGI Zeitgeist Tracker. The four tests, the figure it scores for AI doing the work of improving AI, and the table of what the companies have said, each row linked to its source.

The Belt Equation, step C1: “A non-human mind produces knowledge beyond us”. The rule for when an idea counts as the machine’s rather than the people’s who set up the search.

External Works

Marina Favaro and Jack Clark — When AI builds itself (Anthropic, 2026). The 80 percent figure, what it measures, and the company’s account of what its people now do.

Anthropic — “Measurements for understanding the pace of AI development inside frontier labs” (2026). The prototype index behind the 26 percent figure, with its method and its limits.

Anthropic — “The AI expectations gap” (October 16th, 2026). The staff poll behind the 51 percent figure.