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

A lab reports its own model's role in building the next one

For years, recursive self-improvement was mostly a thought experiment: an AI good enough at AI research to design a better version of itself, which then designs an even better one. Over the past four months, Anthropic has started treating it as something it can measure inside its own company. The company's position, stated plainly, is that it hands a growing share of AI development to AI systems, and that this is speeding up its work1. It defines the endpoint, recursive self-improvement, as a system that can fully and autonomously design and build its own successor. It says that point has not been reached and is not inevitable, but could arrive before most institutions are ready1.

The framing first appeared in a June 4 essay titled "When AI Builds Itself"1115. It has since been extended with internal metrics, follow-up papers and a September update. The most quoted new number is that, as of August, Claude "leads" 26% of Anthropic's research and development. Anthropic says this means the model can finish most of a task end to end from a high-level prompt while humans still supervise10. It also says Claude assists on more than 90% of R&D work6. Outlets agree on the 26% figure but describe the baseline differently. One report says Claude led none of the work in February10. A summary from the governance group GovAI puts it at 1% in March7. Either way, the claim is a rise from nearly zero to a quarter in about six months.

What the internal data shows, and what it doesn't

Anthropic's case stands on several separate kinds of evidence. On engineering output, it says engineers now ship about eight times as much code per quarter as they did on average from 2021 to 20251. Coverage of the June essay reported that Claude was writing more than 80% of the code merged into Anthropic's production systems16. Anthropic also says human code review has become the new bottleneck19.

On reliability, the company says staff step in, correct or take over from Claude less and less often. On its most open-ended task tier, Claude's success rate reached 76% in May 2026, up 50 percentage points in six months1. A September update charted success rates across all four task types converging at roughly 88 to 92 percent1.

The experimental results are the most striking. In a recurring internal test, Claude is asked to make a small model's training code run as fast as possible without breaking correctness checks. Claude Opus 4 averaged about a 3x speedup in May 2025, and Claude Mythos Preview reached about 52x by April 2026. Anthropic says a skilled human would need four to eight hours to get to 4x1. In a weak-to-strong supervision project, Claude-powered agents closed 97% of the gap between a performance floor and ceiling over 800 cumulative hours and about $18,000 in compute. Two human researchers working for about a week closed roughly 23%1.

Anthropic's own caveats deserve as much weight as the headline numbers. The weak-to-strong result did not carry over cleanly to production-scale models, and humans still chose the problem and wrote the scoring rubric1. The company says it expects that the median self-reported productivity gain of about 4x, from a March poll of 130 research staff, overstated the real effect1. It also concedes that Claude still falls well short when it has to use judgment to choose goals1. That is exactly the step that separates a very productive assistant from a system that can direct its own successor.

The "research taste" question

Across the coverage, the open question is whether AI can develop research taste: the judgment to pick problems worth solving and to know which results to trust16. Anthropic offers early, narrow evidence that this is improving. It took moments from real research sessions where a human had gone down a less productive path. In those moments, Opus 4.5 suggested a better next step than the human 51% of the time in November 2025, and Mythos Preview did so 64% of the time in April 20261. The company notes that it picked these moments specifically because the human's choice had room for improvement, so this is not a fair head-to-head comparison1.

Outside benchmarks are now being used to sharpen the timeline. Anthropic co-founder Jack Clark wrote in his Import AI newsletter that he now thinks full recursive self-improvement is more likely to become possible in 2027 than in 20283. He pointed to TasteVal, a benchmark from P-Zero Research, which reports that AI's "experimental research taste" has doubled about every three months since December 2025. On that benchmark, Anthropic's Opus 5.5 beat a baseline of experts from labs including OpenAI and Google DeepMind at about one-thirtieth of the cost per attempt3. Reporting from Korea added a caveat from P-Zero's own researchers: the tasks were easy and quick to grade, which makes it easier for developers to tune against them, so the measured rate of improvement may be inflated8. The benchmark also covers running experiments and interpreting results, not choosing which problem to work on3.

One paper from Anthropic's fellows program points the same way. Automated alignment researchers built on Claude Opus 4.8 improved a target model on all 10 misalignment benchmarks without hurting its overall performance4. The best automated method beat what experienced human researchers proposed in about six hours, at roughly $4 per hour of inference compared with $150 per hour for human staff4.

The open-weight angle: replication without a datacenter

The least-noticed thread for open-source AI is a separate Anthropic Frontier Red Team study published October 125. It asked whether small models could copy themselves across cheap internet-of-things hardware. Anthropic concluded that small models may be only months away from self-replicating on devices like smart ovens and off-brand Android TV boxes2. Every model tested was open-weight. A 9B model got inference running on 8 of the 9 hardware types tested2. In a simplified network, GLM-5.2, from China's Zhipu AI, spread to all eight emulated nodes in four of five runs8. Copies of Qwen3.5-9B that inherited notes from earlier copies raised their setup success from 45% to between 61% and 84%5.

Coverage agrees on the limits. No such worm has been found in the wild5. Device security was deliberately weakened, and GLM-5.2 took instructions from an external server8. Spread would be slow, with doubling times measured in days2. Still, the study links the self-improvement story to a concern specific to open weights. A system with no central host has no central off switch2. And a version of the study's "learning across copies" result shows up in a model whose weights anyone can download.

Outside Anthropic, the open research literature is building its own loops. A July survey on arXiv describes A-Evolve-Training, a system that it says ran the entire post-training loop of a 30B model for weeks with no human involved and nearly matched the top human leaderboard entry13. The same survey argues that the bottleneck in setting research direction and the bottleneck in verifying results are the same problem13.

Agreement, skepticism and the policy push

The coverage splits mainly over motive, not data. Anthropic has said it would be good for the world to have the option to slow or briefly pause frontier development11. Leaders from both Anthropic and OpenAI have voiced alarm, with an OpenAI alignment researcher calling the race toward recursive self-improvement extremely dangerous. Clark and OpenAI's Jakub Pachocki joined Geoffrey Hinton and Yoshua Bengio on a September 28 paper calling for limits on how fast capabilities advance and for outside auditors inside labs7. That paper estimates that progress could speed up about tenfold within 1.5 years if AI R&D were fully automated, under assumptions it states openly7. Critics see a strategic side to this. Georgia Tech's Mark Riedl described labs piling onto a "hype train," and another skeptic argued that more AI-written code shows the tools are useful, not that they represent a great leap11.

Our reading: both camps have part of it right, but the skeptics are answering a weaker claim than the one Anthropic now makes. Lines of code really are a poor proxy. The 52x speedup, the weak-to-strong result and the falling intervention rates measure something different: AI doing research work, not only typing it1. The serious weakness is that most of this data is self-reported and cannot be checked from outside. Anthropic seems to recognize this. It has urged other labs to publish comparable metrics using a public methodology10.

That call for public methodology may matter more than any single number. If recursive self-improvement is arriving, outsiders will need shared metrics, open evaluations and reproducible benchmarks to see it coming. They should not have to rely on newsletters and company blog posts alone.