Under 1 percent in February 2026. Twenty-six percent in August. Six months, one lab, and a chart that will lose its footnote the moment it lands in a strategy deck.
What moved is not a score for the Claude line. It is the share of Anthropic’s own research tasks graded as work the model leads. Claude is not directing a quarter of the world’s AI research, and it is not running the lab with the lights off. Anthropic graded a slice of its own work. On that scale, “leads” still means a person is supervising.
What the 26 percent actually counts
Anthropic’s institute page, “Measuring and Reporting AI R&D Automation and Oversight at Anthropic, ” says that as of August 2026 Claude leads 26 percent of Anthropic’s AI R&D work. A chart on the same page uses a tighter phrase: 26 percent of Anthropic’s model R&D tasks, up from under 1 percent in February 2026. The page is at anthropic.com/institute/measuring-pace-of-ai-development. Those two wordings are close. They have not been shown to share a denominator, so both should stay in view.
Either way, the base is Anthropic. Not the industry. A lab scoring itself cannot hand you a benchmark for every research group in the field, however clean the chart looks. There is no generic “Anthropic Index” behind the figure. If a slide uses that name, ask which page it came from.
It is also not the Anthropic Economic Index, the separate product Anthropic uses to track AI’s effects on the economy, last updated June 26, 2026. Fold those two together and a lab-internal automation snapshot turns into a ghost statistic about the labor market.
Leading a task is not setting the agenda
Anthropic places the work on an automation scale it attributes to Epoch AI. Three rungs are enough. AL3, “AI collaborates, ” means large chunks of work under close human direction. AL4, “AI leads, ” is the 26 percent: the model can complete most of a task end to end from a high-level prompt, while a human supervises. AL5 is fully autonomous work with no human in the loop. Anthropic says Claude is not operating at that level for any measured subset of this work.
Picture a strong engineer who takes a brief, runs the experiment, and brings back a result. Not a research director who decides what the next model should be. People still set the goals. The model executes and tests inside that frame. “Leads” does not mean Claude is picking the direction of the research program.
Step back one level and the picture changes. The share of work at or above “AI collaborates” is above 90 percent. Most of the measured R&D already has Claude in it. The newer, smaller slice is the work where Claude runs the task and the human supervises, rather than the reverse.
If you are weighing AI automation for sales, a finance close, or a support queue, 26 percent is not a target. AI for business does not get more serious by borrowing a frontier lab’s internal ratio. Your work is not Anthropic’s model-training stack. What does transfer is the constraint, and later the control design around AI agents: a high-level brief, a bounded task, and a supervisor who still owns the goal. Handing an agent a vague objective and walking away does not reproduce what Anthropic measured.
A steep climb, measured with a soft ruler
Under 1 percent to 26 percent is the fact worth sitting with. The gap of more than 25 percentage points is only the difference between Anthropic’s two published points, not a separate growth rate the lab certified. Inside a frontier lab, the mix of who does the work can move faster than a planning cycle. That is a change in measured task share, graded AL4. It is not a 26-fold capability jump for a model card.
The ruler is softer than the headline. Anthropic notes that labs do not share a common methodology, and that a developer using its own models as judge can repeat the evaluated model’s mistakes. That is a structural problem, not a footnote for specialists.
Anthropic’s measuring page states the shares. A writeup on finance.biggo.com of what it dates as Anthropic’s September 17, 2026 disclosure describes the sample, and attributes that design to Anthropic. In July 2026, Anthropic randomly sampled 20 percent of employees each week from departments involved in model R&D. A Claude research agent reviewed Slack logs and internal documents. After four weeks, the lab had collected about 15, 000 granular model R&D tasks. The set spanned model training, product development, reinforcement learning, evaluation systems, and engineering troubleshooting, so the 26 percent is not the automation rate of one narrow task type. Claude agents assessed how each task was done. Another Claude model assigned the automation level. Employees scored the same tasks without seeing the model’s scores.
The reported agreement rates are why this is directional, not a KPI. Exact match between the model judge and human scorers was 59 percent. Exact agreement between humans was 35 percent, which says the categories are hard even for people doing the work. Agreement within one tier reached 97 percent. That last figure is the strongest of the three, and it also means many tasks called Claude-led may sit right beside collaboration, not clearly above it.
Thirty thousand agents, and a very small block rate
The operational counterpart is more useful to a company than the percentage. On Anthropic’s most-used internal platform (not a count of every agent the company runs), about 30, 000 agents were doing research and engineering work at any one time.
Automated monitors inspect 100 percent of those agents’ actions before execution, usually within seconds. They are aimed at fast, hard-to-undo harm, such as copying model weights out. Offline monitors ingest 100 percent of the actions after the fact, aimed at slower and more reversible problems, including signs of misalignment. That “100 percent” is a machine check. It is not a room of people watching 30, 000 agents.
Over August 2026, Anthropic analyzed over a billion research-and-engineering agent decisions. The online monitor blocked 0.002 percent of them. A rate of 0.002 percent is 1 in 50, 000. Anthropic has not published the raw blocked count beside that percentage, so use the rate, not an invented headcount of stops. It is a block rate. It is not an error rate, and it is not a measured rate of dangerous behavior. A quiet monitor can mean the agents stayed inside the rules, the rules were narrow, or both.
Blocking and flagging are different cuts, on different bases. Do not add them.
- Online blocks: 0.002 percent of over a billion August decisions, stopped before the action ran. That is 1 in 50, 000. It is not a weekly count.
- Weekly flags: roughly 100, 000 transcripts a week. Most are resolved by classifiers and an initial oversight pass. About 50 a week are escalated to human review. Anthropic also says people review blocked actions within a week. Those two human steps sit next to each other in the disclosure. They are not published as one counted queue, so you cannot tell whether the 50 escalations are the blocked actions, a slice of the flags, or a different pile. They are not a breakdown of the 0.002 percent.
- Offline flags: roughly one to two transcripts in every thousand, after the fact. Different base. Do not add this rate to either number above.
Anthropic says these measurements also appeared in its August 2026 risk report, at anthropic.com/aug-2026-risk-report. It says METR has independently red-teamed the offline monitoring platform, and that it is setting up external third-party evaluators. The point is pace: measure the work, then let someone outside the lab check the measurement. Anthropic has said any frontier lab could publish measures like these with a public methodology, and that another lab’s models or a third party could check them, with limits where the data is competitively sensitive. It has also said such measures could later trigger stronger rules, including a fixed testing window before a new model is used for further AI R&D. Whether that becomes policy is open. Publishing the measure is not the same as a rule other labs must follow.
A safety share that will be quoted as a budget
The same release includes another ratio that will be quoted as if it were a standing budget. It does not change what the 26 percent means. On the measuring page, Anthropic says it examined all of its compute from July 13 to July 20, 2026, and sorted uses among more powerful models, serving customers, and safety work. That safety work included reviews of the model’s written reasoning, training test models built to misbehave so the lab can study misalignment, and running deployment evaluations.
The finance.biggo.com account of the September 17, 2026 disclosure, attributing the figures to Anthropic, put safety at about 6 percent of AI R&D compute, and about 12 percent when the cut was limited to AI-driven AI R&D. In that accounting, safety meant alignment, interpretability, and safety red-teaming. The same account says Anthropic called the estimate extremely conservative, because any compute that benefited both safety and capability was left out of the safety tally. Safety work is also heavier on human experiments than on raw compute, so a compute share understates the labor behind it. Treat those two percentages as a one-week snapshot with a strict exclusion rule, not as the share of the research program the lab “spends on safety.”
Questions worth asking before you repeat the number
Is 26 percent of AI research now Claude-led?
No. As of August 2026, Anthropic says Claude leads 26 percent of its own AI R&D work, while a chart on the same page puts 26 percent on its model R&D tasks. One lab, not the industry, and not proven to be one denominator. It is also not the Anthropic Economic Index, last updated June 26, 2026, which tracks effects on the economy rather than who leads the lab’s research tasks.
Does “leads” mean the model works alone?
No. It is automation level 4: the model can finish most of a task from a high-level prompt while a human supervises. Anthropic says no measured subset reached full autonomy, the level with no human in the loop.
How fast did this change?
Anthropic’s chart shows Claude-led model R&D tasks under 1 percent in February 2026 and at 26 percent in August 2026. Over that period, the share of measured work at or above “AI collaborates” was above 90 percent.
How solid is the grading?
Directional, not a KPI. Anthropic warns that labs lack a shared method and that a judge built from its own models can repeat the model’s errors. Reported exact agreement with human scorers was 59 percent. Humans agreed with each other exactly only 35 percent of the time, and agreement within one tier was 97 percent.
What does the oversight look like in practice?
On the most-used internal platform, about 30, 000 research and engineering agents run at any one time, and in August 2026 Anthropic reviewed over a billion of their decisions. The online monitor blocked 0.002 percent of them, which is 1 in 50, 000. That is a block rate, not a danger rate, and it is not the same figure as the weekly flags or the offline flag rate.
The number will travel without its footnote. When it does, the correction is short. Claude’s share of leading Anthropic’s research work jumped in half a year. It did not take the wheel. For anyone building with AI agents, ask whether the automation has a supervisor, a trace, and a person who still owns the goal. Whether 26 percent sounds large is the wrong test.