Anthropic’s AI Research May Be Accelerating Nearly Twice as Fast, Independent Study Finds

November 5, 2026 — An independent assessment of Anthropic’s AI research process suggests that the company’s progress on certain model-capability benchmarks has accelerated sharply. But the findings do not establish that artificial intelligence can independently develop its successors, and the researchers warn that some of the most informative evidence remains confidential.

What the researchers actually measured

The nonprofit research organization METR reported that Anthropic’s internal model benchmark scores have improved at an estimated 1.9 times the rate observed before October 2025. Its statistical 95% confidence interval spans 1.5 to 2.8 times the earlier rate. That is an estimate of the pace of improvement, not a claim that today’s models are 1.9 times as intelligent or that research productivity has doubled in every department.

METR’s team conducted a high-access assessment between August 12 and October 29, examining internal model results, training-process information, staff interviews and other research evidence. The organization said AI assistance likely explains a majority of the acceleration, while acknowledging substantial uncertainty about the relative roles of other changes.

Why the result matters

AI developers increasingly use their own systems for coding, experiment design and analysis. If better models help researchers build the next generation more quickly, capability gains could reinforce one another. That prospect has implications for competitive investment, independent safety evaluations and the ability of oversight processes to keep pace.

However, the assessment found no evidence that AI systems already make major, fully autonomous contributions to the most important stages of model training. Researchers described AI assistance as the primary mechanism. They also concluded that the faster benchmark progress was not driven by faster growth in training compute, though details of that evidence were redacted.

The important limits

The study covers one developer, relies partly on internal and unreleased-model measurements, and includes subjective judgments about causation. Anthropic was allowed to redact non-public information, and METR said some of the material most relevant to its conclusions could not be disclosed. Consequently, outside readers cannot independently reconstruct every step of the assessment.

Benchmark improvements also do not automatically translate into equivalent gains in real-world reliability, scientific discovery or safety. METR’s report describes uncertainty around how to measure progress and whether future bottlenecks will emerge. Its warning about possible further acceleration is a risk assessment, not a prediction that runaway self-improvement is inevitable.

What to watch next

The most consequential follow-up will be whether similar results appear across other AI laboratories, whether independent researchers can examine more evidence, and whether evaluation and governance methods keep pace with increasingly capable systems. For now, the strongest supported conclusion is narrower but still notable: within Anthropic, selected model-capability metrics appear to be advancing faster, and AI-assisted research may be an important reason.

Sources: METR’s November 5 independent assessment; Reuters, November 5. Editorial note: This report summarizes an independent evaluation and distinguishes measured benchmark trends from broader claims about autonomous AI development.

Please follow and like us: