Anthropic has released an internal estimate of how much of its own AI research and development work is now being led by Claude. As of August 2026, the company said Claude led 26% of Anthropic’s AI R&D work, up from less than 1% in February.
The number sits inside a larger question the company is openly wrestling with: if AI starts helping build the next generation of AI, what role is left for humans in the loop? Anthropic frames the endpoint as “recursive self-improvement,” meaning a model that can fully and autonomously build its successor. By Anthropic’s own accounting, that point has not arrived. No measured work had reached the fully autonomous level as of August 2026.
Jack Clark’s timeline and the caveats around it
On May 4, 2026, Anthropic co-founder Jack Clark used his Import AI newsletter to put odds on what he described as the day the last human leaves the loop. He wrote that the probability of an AI system powerful enough to build its successor without any human involvement appearing by the end of 2028 was 60% or higher.
For 2027, he gave the probability as 30%. Clark said he marked that estimate down because AI research may still require “creativity and heterodox insight,” and systems had not yet shown that ability in any major way.
Five weeks later, on June 8, Clark wrote again that he had spent the previous months inside Anthropic gathering evidence that a smaller version of the same idea had already begun: compound acceleration in the lab’s own productivity. He then answered his own question bluntly: “Do any of these settle the matter? No.”
On September 15, in an interview with NPR, Clark made another limitation explicit: “We don’t have a third party validating what we say.”
How Anthropic produced the 26% figure
The September 17 Anthropic post carrying the number was written by Marina Favaro and Phillie Wright, with Clark listed as research lead. Its headline result was simple: by August 2026, Claude led 26% of Anthropic’s AI R&D work, versus less than 1% in February.
The counting itself was done by Claude rather than by human reviewers. During each week of July 2026, Anthropic randomly sampled one-fifth of employees from every department involved in the model research loop. A Claude agent then reviewed each sampled employee’s week by reading Slack messages and internal documents and writing down the tasks that person had worked on.
The weighting scheme worked at the level of one person-week. Each employee received one unit of weight, split evenly across that week’s tasks. If a person had four tasks, each task received one-quarter of a unit. If the person had 10 tasks, each received one-tenth. Anthropic described this as a rough approximation, but said it performs reasonably at the aggregate level because task categories repeated across many employees accumulate the most weight.
After applying that process across all sampled employees in each week of July, Anthropic said it ended up with a flat list of roughly 15,000 fine-grained tasks.
Those tasks were then placed into a task tree with 542 nodes at different depths, including 378 leaf nodes at the finest level of granularity. Anthropic said the tree was frozen so that each month’s rating would be applied to the same basket of work. The leaves read like a week inside a frontier lab: evaluation platform bug diagnosis and repair, reinforcement learning sandbox egress and network policy, and service incident postmortems.
The six-level scale and where humans still sit
Anthropic used Epoch AI’s six-level automation scale. The scale is defined by the human role in the task, not by the outer limit of what a model might be able to do under ideal conditions.
At the “assistive” level, AI helps with research and polishing. At the “collaborative” level, AI completes large chunks of work, but humans make the decisions. At the “leads” level, AI handles most of the task from a high-level instruction, while humans supervise, correct, and approve. At the “fully autonomous” level, AI works independently and no human remains in the loop.
Anthropic said “leads” is level four on the scale. As of August 2026, no measured work had reached level five, which the company described as fully autonomous. It also reported that more than 90% of work fell into “collaborative” or higher.
The company stressed that the ratings describe how work is usually completed inside the lab, not what Claude could do in a best-effort demonstration.
A concrete example: fixing a broken overnight data pipeline
Anthropic’s page included a hypothetical example to make the levels more tangible. An overnight data pipeline breaks and must be fixed before the next run.
At the collaborative level, an engineer gives Claude the logs, possibly along with a guess about the cause. When a new issue appears, Claude stops and the engineer decides whether to work around it or fix it properly. After tests pass, the engineer may read the changes line by line, rerun the pipeline, and deploy it.
At the leads level, the engineer only hands Claude the alert. Claude then identifies the failing stage and the cause, writes and tests a fix, reruns the pipeline on a copy of the data, compares the output with the last known good run, and writes up what went wrong. Claude does not deploy. It pings the engineer, who reads the report, reviews the changes, asks a few questions if needed, and decides whether to ship that night or wait.
At the higher level that Anthropic says it has not reached, Claude would be trusted to monitor, fix, and deploy, with humans stepping in only if they wanted to.
The method has visible boundary problems
Anthropic also published evidence that the rating system has soft edges. Domain leads rated their own areas without seeing what the reviewer had seen. Reviewer-human agreement on the exact level was 59%. Human-human agreement on the exact level was 35%. Reviewer-human agreement rose to 97% when the comparison allowed a difference of no more than one level.
The lab said plainly that there is room for disagreement in boundary cases, especially over where “collaborative” ends and “leads” begins. That matters because a task moving from one side of that line to the other changes the 26% total.
Anthropic’s chart showed the monthly reading rising from less than 1% in February to 26% in August, with each month rated against the same July work basket. The chart also included a 90% measurement interval and other automation levels, though those details were not redrawn in the article summarized here.
The result is a single lab-wide number. It is not broken down by work type, so it does not show which tasks make up the 26%.
Engineering work and research judgment are not the same thing
Clark drew that distinction in May. He wrote that AI can now automate a large amount of AI engineering work, perhaps all of it, but it remains unclear how much of AI research it can automate.
Epoch’s authors make a similar split. One category is decision-making: what to do, what to continue, and where to allocate compute. In Epoch’s own example ratings, “deciding what to release” and “setting team direction” receive a score of 1, meaning marginal help, while “writing research code” receives a score of 4, the level Anthropic labels as leads. Where Anthropic’s 26% sits between those poles is exactly what the aggregate number cannot answer.
Examples from inside the lab point more toward engineering. In one article co-written by Clark, Claude was said to have delivered more than 800 fixes in April, cutting one class of API errors by a factor of 1,000. The supervising engineer estimated that the workload would have taken humans four years, and one employee reportedly had not written code for about five months.
Another employee wrote that when everything is running smoothly, what they do can feel unimportant; when everything breaks, “I no longer know what I’ve been doing all this time.”
Anthropic’s August risk report gave one reason for keeping humans in the loop: trust. Researchers could have used very large amounts of AI labor on valuable work, the report said, but chose more moderate use because there were steps they did not trust the model to get right.
The same report drew a second conclusion from that fact. Anthropic said it had every reason to replace researchers with models, but had not done so, and took that as a sign that models still could not replace an entire research team. The report also said that interpretation had only been partially tested, because the lab had never spent five times the cost of a researcher on tasks where the model failed in order to see whether money could close the gap.
The article also pointed to OpenAI. As of mid-August, OpenAI’s research organization was running 3.1 agent workdays per human day, and users at the 90th percentile were spending more than $7,000 a day on tokens. Even there, people were still setting research priorities, judging which ideas and results were worth pursuing, and deciding whether to scale, pause, or deploy.
The line drawn in the piece is clear: engineering is the domain Clark has already handed to machines. Research remains the open question.
The 26% figure measures handoff share, not the speed of progress
Anthropic’s June statement was about speed — compound acceleration in the lab’s productivity. The 26% figure published on September 17 measures something else: the share of work that has been handed off to AI leadership.
On September 30, Chris Painter, president of Model Evaluation and Threat Research, or METR, entered the 26% figure into the Senate record in written testimony for a hearing titled “Rogue AI: Securing the Homeland Against AI Agent Attacks.” He wrote later in that testimony that when companies use AI to automate AI research, the feedback loop “may lead to self-sustaining acceleration in capabilities.”
Anthropic’s own rules, however, say the relevant trigger is rate, not share. The company defines “rapid acceleration” as a doubling in the rate of progress — one year of work producing what would otherwise have taken two years. The rules also say this is not the same as doubling researcher productivity, because doubling inputs does not necessarily double the rate of progress.
According to Anthropic’s interpretation in its August risk report, research was moving materially faster than it would have without AI assistance during the period up to July 15, but had not reached 2x. The lab said it was uncertain about that conclusion and that the measurement was difficult. It also said early signs of acceleration had appeared and that the threshold could be crossed within the next year.
Five days after Anthropic published the 26% figure, the system card for Claude Opus 5.5 said internal measurement had not found a sustained, AI-attributable 2x acceleration.
METR then published a summary of its pre-release evaluation of Claude Opus 5.5 on September 22. In that summary, METR relayed an estimate from another METR team with broader access: overall capability acceleration from AI was about 1.5x, with roughly a 30% probability of reaching 2x. That team could not share its evidence with the summary authors, and the estimate did not specify the time period it covered.
So the three numbers are not measuring the same thing. The September 17 figure measures handoff. Anthropic’s August report and the later system card estimate the rate of progress. None of them provide both measures over the same time period for the same class of work.
What the number does and does not say
By Anthropic’s own accounting, Claude led 26% of the lab’s AI research work as of August 2026, and that share rose for six straight months on the same basket of July work. The result is Anthropic’s measure of handoff, gathered by Claude and checked against ratings from employees responsible for the work.
It does not, by itself, say how close the lab is to a closed loop. Anthropic’s rules focus on speed, and the lab’s own estimates remain below the line those rules draw. The unresolved issue is what exactly sits inside the 26%.
If the share is mostly engineering, then it measures the part Clark already viewed in May as within reach. If it includes decision-making and research judgment, then it measures the part he left open.
The article leans toward the first reading. The rating system places decision-making low on the scale. Anthropic’s own examples leave release decisions to humans. Researchers still retain the trust gate. But the author also notes that this is an interpretation, not a measurement. A task-type breakdown would be needed to test it.
Anthropic’s description of the highest level includes a clause about humans that is hard to miss: no human intervention is needed unless humans want to participate. One level below that — the level closer to Anthropic’s current reality — the last person in the loop is an engineer holding a report and deciding whether to ship that night.
The piece was written by Robonaissance and translated by TechFlow.
