# Anthropic's AI r&d pace doubles since late 2025, external evaluators find > Anthropic has roughly doubled its AI capability improvement rate since October 2025, with benchmark growth surging to 1.9 times the historical pace. METR estimates about 80% of that acceleration comes from the company's own AI assisting its R&D. - Type: AI news - Page: https://completeaitraining.com/news/anthropics-ai-rd-pace-doubles-since-late-2025-external/ - Published: 2026-11-06 - Updated: 2026-11-06 - Author: Jeroen Erne - Jobs: Education, Executives and strategy, Government, Human resources, Management, Science and research - Reading time: 3 min A new independent assessment from METR finds that Anthropic has roughly doubled its rate of AI capability improvement since October 2025, and a majority of that acceleration is likely due to the company's extensive use of AI to conduct its own R&D. The evaluation, conducted with high-access from August 12 to October 29, 2026, warns that the lack of clear bottlenecks could lead to even faster - and riskier - compounding progress. The report, published November 5, 2026, offers a rare, partially-redacted look inside a frontier AI lab's development pipeline. While METR staff were given unprecedented access to internal data, Anthropic exercised its right to redact a majority of the most informative evidence, a move the company said was "aimed at removing confidential information that could accelerate other AI developers." #### Benchmark growth has surged to 1.9x the historical rate METR's central finding rests on an analysis of AECI, Anthropic's version of the Epoch Capabilities Index. The data shows that since October 2025, benchmark scores for Anthropic's models have grown at an estimated 1.9 times their pre-October 2025 rate \[95% confidence interval: 1.5-2.8x\]. In practical terms, the company achieved roughly two years of historical progress in a single year. The report notes this acceleration is greater than the 1.7x figure published in a recent system card, primarily because METR included data from newer, unreleased models. METR is confident this surge was not driven by an acceleration in training compute scaling. "We received convincing evidence that the acceleration was not due to an acceleration in the scale of training compute," the report states, though the supporting details were redacted. #### AI use is a primary driver, but full autonomy is not yet here Interviews with Anthropic researchers and internal surveys point to AI as the primary engine of the speedup. Staff self-reported dramatic uplift from using AI systems, with the most recent Slack polls showing estimated speedups reaching 12x by October 2026. METR's overall impression from interviews was a range of estimates centered around a 2x acceleration. The assessment makes a critical distinction: AI is accelerating R&D primarily through assistance, not fully autonomous contributions. "We understand that AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process," the report says. However, it warns that researchers described "signs of life" suggesting such autonomous contributions could occur soon. METR's subjective estimate is that a majority - perhaps 80% - of the current acceleration is attributable to AI, with non-AI factors playing a larger role in the initial acceleration in late 2025. The report acknowledges substantial uncertainty in this judgment. #### No clear bottlenecks to even faster progress One of the assessment's most concerning conclusions is what it did not find. METR did not identify any convincing bottlenecks in algorithmic progress, training compute, experiment compute, or data that would prevent further acceleration in the next 12 months. "Because of this, we believe there is a significant risk that further capabilities progress could lead to even faster acceleration over the coming year," the report states. This finding comes with a significant caveat: METR's understanding of potential bottlenecks changed several times as they received more detailed information, and they are not confident in their current assessment. Anthropic also redacted METR's analysis of several specific bottlenecks. #### Heavy redactions obscure the full picture The public report is a heavily redacted summary of METR's roughly 100-page full report. Key findings that most informed METR's risk conclusions were cut. The redactions covered detailed acceleration estimates for specific parts of the model training pipeline, a breakdown of progress attributable to compute versus algorithms, and forecasts for future models. "As a consequence, the public has to rely on METR's judgment much more than would be ideal, given the weight of the questions answered here," the report acknowledges. METR notes that much of the redacted evidence had multiple plausible interpretations, and both METR and Anthropic researchers have significant uncertainty about how to synthesize it. The assessment process itself was unusual. METR staff worked on Anthropic devices for security reasons and used free tokens provided by Anthropic for AI-assisted data analysis. While METR did not take payment, the arrangement required negotiation over what data could be accessed. Anthropic also canceled a planned compute allocation verification component midway through the project. #### Why this matters for executives and strategy For leaders tracking AI risk and competitive dynamics, this report provides a concrete data point that AI-driven R&D acceleration is not a theoretical future scenario - it is measurable and happening now inside a major lab. The 1.9x acceleration figure, even with its wide confidence interval, signals a structural shift in development velocity. The absence of identified bottlenecks suggests this trend may not plateau on its own. Organizations building risk assessment frameworks or responsible scaling policies should treat compounding AI R&D acceleration as a near-term operational concern, not a speculative one. The heavy redactions also underscore a growing tension: the evidence most critical for public risk evaluation is precisely the information companies are most reluctant to share. --- This is the Markdown form of https://completeaitraining.com/news/anthropics-ai-rd-pace-doubles-since-late-2025-external/. The whole site is mapped at https://completeaitraining.com/llms.txt.