Anthropic's 26% figure grades Claude's R&D tasks by how much human steering they took
The lab says Claude leads 26% of its model research and collaborates on about 90%, with both tiers defined by how closely a human directs each task, and it wants rival labs publishing the same measure.
The Engineer · Build desk
What happened
- About 90% of the company's research and development is now done in collaboration with Claude, according to the same announcement.
- The share Claude leads was none in February and had reached a quarter of the work by August, on the company's own account of its progress.
- The company framed the metrics as a way to gauge how close leading labs are to recursive self-improvement, a model's ability to autonomously build its successor.
Why it matters
- Who has to decide decision Review capacity becomes the staffing variable for anyone copying this setup: work in the collaborative band is work a human has to direct closely, so each added agent buys supervisor hours as well as throughput.
- What it limits constraint An engineering manager cannot reproduce this number for their own team without first inventing a task taxonomy and a classification rule, which makes cross-company comparison a matter of trust for now.
- What it sets up precedent By asking other developers to publish the same measures on a shared public methodology, Anthropic sets a disclosure format that rivals will now either adopt or decline in public.
- Who's exposed exposure Embedding external third-party evaluators inside the company gives people outside it standing to check safety work from within, and makes those evaluators the first outside readers of numbers like these.
Next story · Product · Sep 18
Anthropic reports Claude leading 26% of its model R&D on a definition it wrote itselfConfirmed3 publishers
A task Claude "leads" is one it can complete most of "end-to-end from a high-level prompt" while a human supervises, in Anthropic's wording [2]. A collaborative task is one where the model does "large chunks of work under close human direction" [5]. The difference between the two tiers is how much steering each task needed, and the company said the model is not yet working completely autonomously [3]. The gap between the two shares is 64 percentage points of research and development where a person is directing the work closely [22].
For the 26% to transfer to another engineering organisation, you would need the task list it divides and the rule that sorts each task into a tier. The Economic Times account of the announcement does not report either, nor a count of the humans supervising the roughly 30,000 agents the company said were doing research and engineering work in August [4][12]. The growth number has the same dependency. A rise from none to 26% between February and August works out to about 4.3 points a month [23].
The other figure people quote alongside this one counts a different unit. The Sequence, opening a series on recursive self-improvement, wrote that Anthropic published an essay in June 2026 stating that as of May, Claude authored more than 80 percent of the code merged into its production codebase [17]. Merged lines and research tasks are not the same denominator, so the two percentages do not stack. The same newsletter recorded OpenAI's disclosure that GPT-5.3-Codex helped debug its own training process and manage parts of its own deployment, and DeepMind's AlphaEvolve turning up algorithmic improvements that land in the infrastructure other models train on [18][19]. "The question is which parts of the job it took, how well it does them, and what checks the work," The Sequence wrote [20].
Until recently there was nothing to instrument. The Sequence noted that for about sixty years, arguing about recursive self-improvement meant arguing in the abstract, because there was no system to point at [21].
Anthropic said that models accelerating their own development could make it "more challenging for humans to understand or control these systems" [13]. Its blog post located the remedy in disclosure: "This means better measuring the development of AI, reporting on it publicly, and giving society an opportunity to decide how to use this information," the company said [8]. It also said oversight measures matter for seeing how often monitoring systems detect agent misbehavior [9].
An Anthropic researcher resigned last week with a warning about the threats the technology poses to humanity [15]. Chief executive Dario Amodei, OpenAI's Sam Altman, Elon Musk and other tech leaders have since supported slowing development, while other tech leaders and President Donald Trump have pushed back [16].
What to watch
- Whether OpenAI or Google DeepMind publish a comparable "led" percentage with a methodology an outsider can check.
- Whether Anthropic releases the task taxonomy and the human headcount behind the 26% and 90% figures.
- Whether the embedded third-party evaluators report how often monitoring systems detect agent misbehavior.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption65
- Hype gap+30
- Incentives70
- Confidence60
Perspective Coverage
6 publishers- Builder
- Builder 39%
- Operator
- Operator 33%
- Investor
- Investor 28%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Claude is leading 26% of Anthropic's model research and development, the company said.
- [2]
Anthropic said the 26% "led" category means Claude can complete most of a given task "end-to-end from a high-level prompt" while still being under human supervision.
- [3]
Anthropic said the model is not yet working completely autonomously.
- [4]
Anthropic said there were approximately 30,000 agents doing research and engineering work as of August.
- [5]
About 90% of Anthropic's research and development is done in "collaboration" with Claude, which the company said means the model can do "large chunks of work under close human direction."
- [6]
The portion of work Claude leads was none in February and had reached a quarter of research and development work by August.
- [7]
Anthropic urged other AI developers to share similar metrics on a regular basis, encouraging the use of a public methodology so the numbers could be compared over time and potentially across labs.
- [8]
Anthropic said in a blog post: "This means better measuring the development of AI, reporting on it publicly, and giving society an opportunity to decide how to use this information."
- [9]
Anthropic said oversight measures are crucial for seeing how often agent misbehavior is detected by monitoring systems.
- [10]
Anthropic recently committed to setting up external third-party evaluators who will be embedded within the company to monitor safety efforts.
- [11]
Anthropic said in an announcement on Thursday that its model Claude is helping the company develop the next, more intelligent version of the model.
- [12]
The Economic Times account of Anthropic's announcement does not report the task list behind the percentages, the rule used to classify tasks, or a count of the human researchers supervising the agents.
- [13]
Anthropic said that models accelerating their own development could make it "more challenging for humans to understand or control these systems."
- [14]
Anthropic argued that sharing these metrics could show how close leading AI labs are to recursive self-improvement, defined as a model's ability to autonomously build its successor; it was unclear from the disclosure how close the company believes it is.
- [15]
An Anthropic researcher resigned last week with a dire warning about the threats the technology poses to humanity.
- [16]
Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, Elon Musk and other tech leaders have since supported the idea of slowing down development, while other tech leaders and President Donald Trump have pushed back.
- [17]
The Sequence wrote that in June 2026 Anthropic published an essay stating that as of May, Claude authored more than 80 percent of the code merged into its production codebase.
- [18]
The Sequence wrote that OpenAI disclosed GPT-5.3-Codex helped debug its own training process and manage parts of its own deployment.
- [19]
The Sequence wrote that DeepMind's AlphaEvolve has been turning up algorithmic improvements that land in the infrastructure other models train on.
- [20]
"The question is which parts of the job it took, how well it does them, and what checks the work," The Sequence wrote.
- [21]
The Sequence wrote that for about sixty years, arguing about recursive self-improvement meant arguing in the abstract, because there was no system to point at.
- [22]
The share of Anthropic's R&D done in collaboration with Claude exceeds the share Claude leads by 64 percentage points.
- [23]
The rise in the share Claude leads averages about 4.3 percentage points per month over the February-to-August period.
Sources
6 independent publishers whose own reporting we read for this story.
- dev.toAgent Oversight Needs Metrics, Not Just Logs
1 article · September 18, 2026
- economictimes.indiatimes.comAnthropic Claude: Anthropic says its model Claude is helping to build the next version of itself - The Economic Times
1 article · September 17, 2026
- mezha.netAnthropic заявила, що Claude виконує 26% робіт над новими технологіями
1 article · September 17, 2026
- the-decoder.comAnthropic wants you to know Claude leads a quarter of its research, but "lead" doesn't mean what you think
2 articles · September 18, 2026
- theneuron.aiAnthropic Says Claude Leads 26% of Its AI Research. The Bigger Story Is Who Sets the Metric
2 articles · September 18, 2026
- thesequence.substack.comThe Sequence Knowledge - Issue 933: When the Factory Starts Building Itself
1 article · September 15, 2026