Anthropic is using Claude to help build better AI, and a new independent assessment suggests that feedback loop is already changing the speed of frontier-model development. After eleven weeks examining internal company data and interviewing researchers, the nonprofit METR estimates that Anthropic’s progress on well-specified tasks is running at roughly twice its earlier pace.
The numbers are striking but not presented as a precise countdown to artificial general intelligence. METR measured Anthropic’s benchmark performance using the company’s version of the Epoch Capabilities Index, finding that scores have risen at approximately 1.9 times the rate seen before October 2025. In practical terms, Anthropic appears to have made about as much benchmark progress in the past year as it made across the preceding two years.
METR’s best judgment is that AI assistance accounts for a substantial share of that acceleration—possibly around 80 percent as of October—although the organization emphasizes that the estimate is highly uncertain. Anthropic employees described using Claude for coding, experimentation, implementation, data analysis, and other research tasks. The immediate effect is less a fully autonomous AI scientist than a workforce able to test more ideas, discard failures faster, and reduce the time between a research concept and an experiment.
That distinction matters. The assessment says AI is not yet making highly significant, fully independent contributions to the most important parts of Anthropic’s model-training pipeline. Researchers did, however, report “signs of life” suggesting that more autonomous contributions could become meaningful soon. METR also found no evidence that the change was simply the result of training-compute growth, and it identified no clear near-term limit from data, experiment capacity, algorithmic progress, or compute.
The investigation covered internal models trained between July and October 2026, including a latest model that cleared more than half of every benchmark on which it was tested. Yet the public findings come with an important transparency problem: Anthropic was allowed to redact confidential material, including detailed evidence about AI’s role in specific parts of development, bottlenecks, future projections, and some interview context. METR says those omissions leave the public relying more heavily on its interpretation than it would prefer.
For Anthropic, the result creates an uncomfortable contrast. The company has positioned itself as a safety-focused frontier lab, while its own systems may be helping compress the timeline between successive generations of increasingly capable models. For the wider industry, the question is no longer only whether AI can automate software engineering or research chores; it is whether those productivity gains will compound quickly enough to outrun the safeguards meant to govern them. METR’s conclusion is cautious, but its warning is direct: if the current acceleration continues and human bottlenecks begin to disappear, capability growth could speed up again.
