In a Nov. 5, 2026 assessment, Model Evaluation and Threat Research said Anthropic is likely improving its models on well-specified tasks at around twice the rate it did from March 2024 to October 2025. METR gave the estimate high uncertainty, saying the true acceleration could be closer to 1.3 times or three times the earlier rate.
What METR reviewed
From Aug. 12 to Oct. 29, 2026, METR staff reviewed internal Anthropic materials, including benchmark data, training-compute metrics, AI speedup analyses, forecasts, and staff interviews. METR said Anthropic provided wider access than expected, but the assessment was not full employee-like access and did not include sampling from non-public models.
The access came with limits. Anthropic had redaction rights over non-public information, and METR said the company redacted much of the evidence that most informed the group’s risk conclusions. Anthropic described the redactions as intended to avoid exposing confidential details that could accelerate other AI developers.
What twice as fast means
METR is not saying Claude can fully automate Anthropic’s research organization. Its estimate refers to the rate of capability gains on well-specified tasks, measured partly through Anthropic’s version of the Epoch Capabilities Index.
METR said internal-model benchmark scores have accelerated 1.9 times since October 2025, with a 95% confidence interval of 1.5 to 2.8 times. In METR’s framing, a 2x acceleration means two years of historical progress occur in one year. The group said Anthropic appears to have made roughly as much benchmark progress over the past year as it made between October 2023 and October 2025.
The estimate comes with caveats. METR said its analysis might overstate acceleration if the benchmarks mostly capture a narrow set of verifiable tasks, if non-AI factors explain more of the gain, or if Anthropic’s earlier growth rate was unusually low. It could also understate the issue if benchmark scores lag behind real-world capability gains or if AI automation compounds further.
AI assistance appears to be a major driver
METR said it was confident the measured acceleration was not caused by faster growth in training compute. It judged, with substantial uncertainty, that a majority of the current acceleration is likely attributable to Anthropic’s extensive use of AI in research and development. The group gave a rough subjective estimate of about 80% for October 2026, while warning that non-AI factors likely played a large role in late 2025.
Anthropic has published related internal measurements. In an Anthropic Institute post, the company said that as of August 2026, Claude was not operating fully autonomously for any measured subset of AI R&D work. But Anthropic said Claude was leading 26% of the company’s AI R&D work, and that the share of work at or above the “AI collaborates” level was above 90%.
Anthropic also said about 30,000 agents were doing research and engineering work at any one time in its most-used internal platform. The company said that platform used online and offline monitoring, including real-time checks before actions and later review of agent activity.
Why the report landed at an awkward moment
The assessment arrives after Anthropic CEO Dario Amodei publicly argued for pacing frontier AI development. In a September essay, Amodei said AI capability development had accelerated in recent months, driven mainly by AI’s growing ability to build the next generation of AI. He called for the industry to slow the pace of model-capability improvements so safety work could keep up.
On Oct. 15, 2026, Anthropic also announced a separate embedded-evaluator arrangement that gives a METR team company hardware, office space, internal tools, and systems roughly comparable to Anthropic’s own risk assessment teams. METR emphasized that the Nov. 5 acceleration assessment was a separate project, not the same employee-like access arrangement.
The tension is not that Anthropic hid the issue entirely. The company has published some of its own measures of AI-led research and development. The problem is that the most decision-relevant evidence is still difficult for outsiders to inspect independently, especially when redactions cover details about specific bottlenecks, forecasts, and parts of the model training pipeline.
METR’s risk conclusion was more cautious than the headline suggests
METR’s companion review of Anthropic’s August 2026 risk report agreed with Anthropic that direct catastrophic risks from models covered through July 15 were low. METR said Anthropic may have crossed the early-warning threshold in its own Responsible Scaling Policy, though researchers on both teams leaned toward it remaining below that threshold as of July 15. METR also said the public risk report understated the level of acceleration taking place.
That distinction is meaningful. Crossing an early-warning threshold is not the same as proving that an AI system can replace researchers or directly cause catastrophic harm. It means the rate of AI-driven progress may be entering a zone where stronger measurement, outside evaluation, and safety controls become harder to postpone.
What this means beyond Anthropic
The report is a snapshot of one company, not a cross-industry measurement. METR’s findings depend partly on Anthropic’s internal metrics, staff interviews, and redacted evidence, so they cannot be treated as a public benchmark for every frontier lab.
Still, the report points to a practical issue for AI buyers, software teams, and policymakers: model releases are no longer the only thing worth tracking. The development process itself is becoming automated. If AI systems accelerate the labs that build them, then questions about monitoring, independent evaluation, compute allocation, and redaction rules become part of the product-risk picture.
I can suggest related Tech Help Canada articles based on the topic you’re reading now.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
