Anthropic says Claude now leads 26% of its internal AI research, up from under 1% in February

Anthropic says its own AI now runs a quarter of its AI research. The company disclosed on Sept. 17 that Claude models lead 26% of its internal AI research and development work. In February that figure was under 1%.

The company grades AI involvement on a scale of automation levels. The level in question is AL4, where the AI leads the work and humans oversee, correct and approve. Anthropic also reported that more than 90% of the R&D tasks it measured involved AI at AL3 or higher as of August 2026.

None of the measured work ran at AL5, fully autonomous with no human involved. Roughly 30,000 AI agents were active on Anthropic's internal platform as of August.

The projection

Anthropic says Claude could lead around 80% of its internal AI R&D at the AL4 level by the end of 2026, if the current trend continues. That is roughly 11 weeks from today, Oct. 15.

That is a company forecast about its own operations. It is conditional, and nobody outside Anthropic has audited the baseline.

Clark moves his timeline

Jack Clark, an Anthropic co-founder and the head of research direction on the September piece, now favors 2027 over 2028 for AI that can fully improve itself. He laid that out in an Oct. 12 post. He gave no updated numerical odds, and he described the forecast as his personal view.

In a May 4 post, Clark put the odds of automated AI research capable of training its own successor at 30% for 2027 and about 60% by the end of 2028. He kept the 2027 number low because research may require creativity and what he called heretical insight, which models had not clearly shown.

He points partly to a new benchmark called TasteVal, built by P-Zero Research. It gives models eight fixed AI research problems. The models design experiments and interpret results, while a separate, fixed coding agent handles implementation.

P-Zero reports that Anthropic's Opus 5.5 led on its measure of experimental compute efficiency. Its paper estimates $282 per Opus 5.5 run versus $9,123 per human run, about one-thirtieth of the cost. The math counts model API charges, GPU time at $2.50 an hour and human pay at $119 an hour.

The human baseline is the best attempt per task from 24 experts recently affiliated with organizations including OpenAI and Google DeepMind.

What the data does not show

The numbers come with limits, and the people behind them state them. P-Zero's authors acknowledge that the tasks have results that are easy to check and that their human baseline excludes top researchers. Opus 5.5 refused one task, which they treated as missing data. A comparison across the other seven tasks produced a similar result.

Clark has been just as direct about the evidence at Anthropic. In a June 8 post he wrote that his gathered evidence suggested a smaller version of the same acceleration had begun. Then he asked, "Do these amount to anything conclusive?" and answered, "No."

On Sept. 15, two days before the 26% figure came out, Clark told NPR: "We do not have a third party to verify what we are saying."

Earlier results also did not always carry over. Anthropic's data shows that experiments in early 2026, in which Claude agents outperformed human researchers on alignment and safety tasks, produced gains that were sometimes not replicable in production models. In April, agents running Claude Opus 4.6 closed 97% of the performance gap on an alignment problem compared with human researchers, in seven days.

That creates an obvious tension. If AI leads more of the research, the humans who would catch those failed transfers are doing less of the hands-on work.

Output is climbing

Anthropic's Oct. 13 engineering post offers a harder number. Code changes merged per day rose from about 2,500 in May to more than 30,000 in September. The company said "coding is no longer the bottleneck," and named source control, builds and continuous integration as the new constraints.

That measures volume, not whether the research is any good. Merged changes say nothing by themselves about whether the models are choosing the right problems.

In a separate test in August, Claude designed protein binders, molecules engineered to latch onto specific biological targets, and posted higher success rates than human experts, according to Anthropic.

The race around it

Anthropic is not alone in aiming at this. OpenAI has set a target of March 2028 for an automated AI researcher, with people guiding the work.

Anthropic's reported results are for its own systems, measured by its own staff, on its own tasks. Those results are the evidence for a claim with large consequences for the labs, for chip demand and for how much human oversight stays in the loop.

Two markers are next. One is whether Anthropic's next disclosure shows the AL4 share moving toward the projected 80%. The other is whether any measured work registers at AL5. Until an outside party checks the numbers, the 26% remains the company's own count.