METR estimates AI-assisted research may double Anthropic's progress speed

The acceleration of AI research

A new assessment by METR suggests that Anthropic may be achieving roughly two years of progress in a single year on clearly defined tasks. This acceleration is attributed to the company's use of AI to assist its own research and development teams. By using AI tools, researchers can build more capable models at a significantly faster pace than previously observed.

The report draws on 11 weeks of access to Anthropic's internal research data and staff interviews. METR calculates that by October 2026, the rate of advancement was approximately double what it was during the period between March 2024 and October 2025. This suggests a potential feedback loop where AI helps researchers create the next generation of AI, which in turn provides even better assistance for future development.

While the findings are significant, METR notes substantial uncertainty regarding the exact cause of this speed. The report acknowledges that hiring and other factors likely played a role in late 2025. However, METR judges that by October 2026, AI probably accounted for most of the current acceleration in the company's development cycle.

From coding to conceptualizing

Interviews with Anthropic researchers reveal a shift in how development work is actually performed. Staff members describe using AI to turn ideas into experiments more quickly, allowing them to explore approaches they previously might not have had the time to try. One researcher noted that the team is no longer bottlenecked on writing code but is instead limited by ideas and compute resources.

Another researcher compared the experience of using AI to having 'postdocs and grad students' available to work toward a larger project. This indicates that AI is currently serving as a high-powered assistant to human researchers rather than acting as a fully autonomous entity. While fully autonomous research is not yet a major source of progress, staff reported seeing early signs of independent AI contributions in specific parts of the process.

METR emphasizes that these gains primarily involve AI assisting human researchers in their workflows. The researchers are using the technology to handle the heavy lifting of experimentation. This allows the human experts to focus on the high-level strategy and conceptual design of the models they are trying to build.

Measuring the progress rate

To determine if models were improving faster, METR examined an index combining results from various tests of what AI systems can do. They looked at Anthropic's version of the Epoch Capabilities Index, which showed benchmark performance improving at an estimated 1.9 times its pre-October 2025 rate. This analysis included recent models that have not yet been released to the public.

METR's broader judgment of a roughly doubled pace incorporates private evidence beyond just these benchmarks. They stated they would not be surprised if the actual speed of progress fell somewhere between 1.3 and three times the historical rate. The report identifies several reasons why these estimates might be too high or too low, including the fact that clearly structured tests might not capture performance on less predictable work.

The assessors also noted that Anthropic's earlier progress was slower than the wider industry frontier, which could make the recent acceleration appear larger. Additionally, because model test results arrive after the research and training are finished, they may lag behind the most recent changes. METR warns that acceleration in these specific indices may not map cleanly to real-world impact.

Data limitations and redactions

A significant portion of the evidence used to reach these conclusions remains private. Anthropic delivered a 100-page full report to METR but redacted a majority of the most informative evidence before the public version was released. The company stated these redactions were aimed at removing confidential information that could help other AI developers.

The withheld material includes detailed estimates of algorithmic progress and specific evidence linking progress to the use of AI. It also includes analysis of specific constraints on future development. METR agreed that the withheld material fits the description of confidential information, though they declined to judge whether the redactions were fully justified.

Because assessors could not freely search internal systems, there is a risk they missed information they did not know existed. METR acknowledged that they may have missed important limits on progress due to this restricted access. They also noted that the public must rely on their judgment more than they consider ideal because of these missing pieces.

The future of AI development

Despite the rapid progress, METR found no clear obstacles related to algorithms, computing power, or data that would stop further rapid advancement over the next 12 months. The assessors' understanding of the development landscape changed repeatedly as more information arrived during the study. They are currently monitoring how these capabilities translate into broader capabilities.

The study also highlights a shift in how companies prioritize their resources. Anthropic scrapped a planned review of how it divided computing power between model capabilities and safety to focus more on risk assessment. This move suggests that as models become more capable, the focus on safety and risk management becomes a primary concern for the developers.

As AI continues to assist in its own creation, the speed of development remains a key variable to watch. If the feedback loop continues to strengthen, the time required to reach new milestones could continue to shrink. The industry is moving toward a model where the primary bottleneck is human creativity and the availability of compute power.

What to watch next

Watch for future reports on how AI-assisted research impacts the safety and alignment of new models. It will be important to see if the 'feedback loop' leads to unpredictable behaviors in models that were developed at accelerated speeds.