Anthropic AI Self-Improvement Study: Claude Now Leads 26% of R&D

What happened

Anthropic has published its most detailed account so far of how much of its own AI development is now done by AI. The company's essay, "When AI builds itself," lays out what it describes as progress toward recursive self-improvement, meaning an AI system that can design and build its own successor without human help. The company says it has not reached that point and that the outcome is not inevitable. It also warns that the moment could arrive before most institutions are ready for it.1

The essay first appeared on June 4 and received a metrics update in mid-September.111 The update's headline figure is that Claude led 26% of Anthropic's research and development work as of August. In Anthropic's usage, "led" means Claude finished most of a task from start to finish while a human supervised. Claude also assisted on more than 90% of tasks, and the company stresses that it still cannot work fully autonomously.6 A summary from GovAI, cited in later coverage, says the 26% share was up from 1% in March.8

Anthropic does not frame this as a curiosity. It says models that speed up their own development could become harder for people to understand or control, and that it is publishing the figures so the world can judge how close recursive self-improvement is.6

The evidence Anthropic is putting forward

The argument draws on two kinds of data: public benchmarks and internal numbers the company had not released before. On the public side, Anthropic says the length of tasks models can complete reliably on their own now doubles about every four months, faster than an earlier pace of every seven months.1 It also notes that AI systems went from reproducing published research results about 20% of the time on CORE-Bench in 2024 to saturating that benchmark roughly fifteen months later.1

The internal numbers are more striking. Anthropic says its engineers now ship about eight times as much code per quarter as they did on average between 2021 and 2025.1 Other coverage of the report says Claude writes more than 80% of the code merged into Anthropic's production systems.16 On one fixed internal test, where models are asked to speed up code that trains a small model, Claude Opus 4 managed about a 3x speedup in May 2025. Claude Mythos Preview reached about 52x by April 2026. Anthropic says a skilled human researcher would need four to eight hours to reach 4x.1

The company also tested research judgment directly. It took real sessions in which researchers had gone down an unproductive path, showed models only the work done before that detour, and asked what they would do next. On that measure, the best model in November 2025 picked a better next step than the human 51% of the time. By April 2026 the figure was 64%.1 Anthropic notes that the moments were chosen because the human's choice was known to be flawed, so this is not a fair head-to-head comparison between models and people.1

The weakest point is the one the company names itself. Claude can now match or beat skilled humans at carrying out a well-specified experiment, but it still falls well short at choosing which goals are worth pursuing. Anthropic calls this the gap between today's systems and one that could design its own successor.1 An arXiv survey of 1,250 papers on self-improvement reaches a related conclusion from outside the company. It argues that research execution can be checked, while setting research direction cannot, and that these may be the same bottleneck.14

A run of supporting studies

The essay is part of a wider stream of work. A paper from an Anthropic fellows-program researcher, dated August 28, describes automated alignment researchers built on Claude Opus 4.8. These agents search the literature, propose training methods, and train target models in short loops. They reportedly improved results on all 10 benchmarks of misaligned behavior they were tested on, without hurting overall performance.4 The paper also says the best automated method beat what experienced human researchers proposed within about six hours on average. It puts the cost at about $4 an hour of inference, compared with $150 an hour for human researchers.4

An earlier demonstration pointed the same way. Agents working on an open AI-safety problem, whether a weaker model can supervise a stronger one, recovered 97% of the performance gap over 800 cumulative agent-hours. Two human researchers working for about a week recovered roughly 23%.1 Anthropic notes that humans still chose the problem and wrote the scoring rubric, and that the result did not carry over cleanly to production-scale models.1

On October 12, Anthropic's Frontier Red Team took the self-improvement question in a darker direction. Its research suggests that small models may be only months away from being able to copy themselves across cheap internet-connected devices such as off-brand Android TV boxes.5 In the tests, a 9B-parameter model got inference running on 8 of 9 hardware types. Anthropic says GLM-5.2 and stronger models could complete end-to-end replication in its simplified evaluations.2 Copies that passed notes to the next generation got better at the job: setup success rose from 45% in the first generation to between 61% and 84% in later ones.5 Korean coverage points out caveats that matter. GLM-5.2 took instructions from an external server, the devices' security had been deliberately weakened, and no fully self-replicating AI exists yet.7

Timelines are getting shorter

On the same day, Anthropic co-founder Jack Clark wrote in his Import AI newsletter that full recursive self-improvement now looks more likely to become possible in 2027 than in 2028.3 He based the change partly on P-Zero Research's TasteVal benchmark, which found that models' "experimental research taste" has doubled roughly every three months since December 2025. On that benchmark, Anthropic's Opus 5.5 beat human experts from labs including OpenAI and Google DeepMind at about a thirtieth of their average cost per attempt.3 P-Zero's own researchers added a caution. The tasks were easy and quick to grade, which makes them easy for developers to target, so the measured rate of improvement may be inflated.7 The test also measures only experiment design and interpretation, not the choice of problem.3

People outside Anthropic are talking about this too. In late September, Clark and OpenAI's Jakub Pachocki joined Geoffrey Hinton, Yoshua Bengio and others on a paper urging preparation for an "intelligence explosion." It proposes limits on how fast capabilities advance and outside auditors working inside labs.8 One scenario in the paper estimates that progress could speed up about tenfold within 1.5 years if AI research were fully automated.8 Researchers at both companies have said publicly that they are worried. One Anthropic safety researcher posted that there is no viable scientific plan yet for the risks of recursively self-improving AI.

Where the coverage diverges

Most reporting agrees on the basic facts. Anthropic is measuring real acceleration and says it is cautious about where that leads. The disagreement is over what the acceleration means. Scientific American gave space to skeptics who see the announcements partly as business strategy, a way to draw regulators to the frontier while Anthropic keeps racing toward it.11 Georgia Tech's Mark Riedl dismissed the industry's "hype train" around recursive self-improvement, and another critic argued that more AI-written code shows the tools are useful, not that a great leap has happened.11

TIME pointed to a contrast in tone. OpenAI has an internal target of March 2028 for fully automating its AI researchers, while Anthropic tends to describe automation as something happening to it, an unavoidable result of continued progress.12 That framing is worth questioning. A company that ships the models doing the work is also choosing to delegate the work.

Anthropic's own caveats deserve attention as well. Its March poll of 130 research staff produced a median estimate of about 4x output uplift from Mythos Preview, but the company says it expects the true figure was somewhat lower.1 Lines of code is a crude productivity measure, and Wikipedia's summary of the field notes that Anthropic itself has said reviewing AI-written code has become a new bottleneck.19

Our reading

The skeptics are right that code volume is not intelligence, and benchmark doubling rates on easy tasks can mislead. But they argue against a weaker claim than the one Anthropic makes. Anthropic's argument does not depend on Claude developing research taste. Even under what it calls the conservative reading, humans who spend their time on the small share of work that sets direction can steer far more output than before, and the acceleration compounds.1 That scenario, compounding efficiency with humans still setting direction, is the one Anthropic says the world is probably heading into.15 The arXiv survey also observes that nothing in the skeptical literature rules it out.14

The real open question is not whether AI is speeding up AI development, since the internal data across these studies points clearly to yes. It is whether the last human role, choosing problems and judging results, holds or gives way. The figures on next-step judgment, the TasteVal trend, and Clark's revised timeline all suggest it is weakening, though each comes with real measurement caveats. Anthropic has said it wants the world to have the option to slow or pause frontier development, and that it will consult governments and rivals on whether a coordinated slowdown could work.11 Whether that option still exists by 2027 may matter more than any single benchmark.