METR Says Anthropic’s AI Research Pace May Have Doubled, With AI Driving Most of the Gain

Anthropic’s model progress appears to have accelerated sharply, and AI assistance in the company’s research may explain most of the increase. That is the central finding of an assessment by the nonprofit METR, based on eleven weeks of access to internal data. The estimate is uncertain, and Anthropic removed much of the evidence METR considered most informative before the report became public.

The question matters because Anthropic’s Responsible Scaling Policy identifies a doubling in the pace of progress due to automated AI research as a warning threshold. METR says Anthropic may have reached that level in July 2026. The underlying concern is a feedback loop: AI systems help researchers build more capable systems, which in turn may accelerate later research. METR describes the possible endpoint as the sudden arrival of highly capable systems that carry significant risks.

Model scores rose faster after October 2025

METR compared Anthropic’s results on the AECI, a composite measure drawing on multiple benchmarks. From March 2024 through October 2025, scores rose by about 15 points per year. Since October 2025, the estimated pace has been 29 points annually. METR says that means the company achieved in about a year what had previously taken two, with the change roughly comparable to the progress from Claude 2.1 to Sonnet 4.5. The analysis also includes three unreleased models whose training ended between July and October 2026.

METR does not attribute the faster gains to a faster increase in training compute. Its estimate is that roughly 80 percent of the acceleration comes from AI being used in Anthropic’s own research. One employee described the change this way: “Now only bottlenecked on ideas & compute, no longer bottlenecked on code.” Researchers also told METR that experiments that once took one or two days to set up had become “~free.”

Informal employee polls in October 2026 put average self-reported productivity with the latest internal model at about twelve times higher. METR cautions that these accounts are loose, unrepresentative evidence and likely to overstate the gains. The reported change across models ranged from 1.2 times with Claude 3.7 Sonnet to 12 times with internal Model 4.

METR’s account does not suggest that AI is conducting Anthropic’s research independently. The current gains mainly come from AI acting as an assistant, though METR sees early signs that models could make more independent contributions. The auditors found no clear bottleneck likely to slow progress over the next twelve months. They therefore see a significant chance that the acceleration will continue to grow.

Estimates carry wide uncertainty

The acceleration estimate could plausibly be as low as 1.3 times or as high as three times, METR says. When results from OpenAI, Google and xAI are included, the estimated industry-wide increase falls to 1.4 times. One possible explanation is that Anthropic was behind competitors in 2025 and has since been catching up, making its own recent trajectory look especially steep.

Anthropic researchers also disagree about what drove the change. One put AI’s contribution to productivity gains at only 15 to 20 percent and questioned the reliability of coworkers’ self-reports. Another considered hiring new employees the largest factor in recent progress. The report’s 80 percent estimate should therefore be read as METR’s assessment, not a settled internal consensus.

METR conducted its review from August 12 to October 29, 2026, with a team of two and a half full-time staff. The nonprofit examined benchmark results from unreleased models, training metrics, employee surveys and internal analyses, and interviewed roughly 15 researchers. It says Anthropic provided unusually broad access. METR took no money from the company and worked on its devices for security reasons. This project was separate from the collaboration Anthropic announced with METR in October, which the nonprofit described as having a different scope. The review followed CEO Dario Amodei’s commitment to bring in outside auditors.

Redactions limit what readers can check

Anthropic had the right to remove non-public material and did so extensively. METR says the company withheld most of the evidence the auditors found most useful, including future-model forecasts, training-step breakdowns and analysis of research bottlenecks. Anthropic said the removals were intended to prevent confidential information from helping other AI developers.

That leaves readers relying more heavily than METR would prefer on the nonprofit’s interpretation. One chart in the report is marked “FAKE DATA,” reflecting the limits imposed on the evidence that could be shown. The restriction makes it harder for outside researchers to assess how strongly the published figures support the conclusions.

July threshold and a canceled review

A related METR review of Anthropic’s August 2026 risk assessment raises the possibility that the threshold was crossed earlier than the company’s public account suggested. Anthropic had described acceleration as “less than a factor of 2.” The unredacted version implies a figure of 1.97, and the company called it plausible that the threshold had been crossed. METR says the warning level may have been reached as early as July 15, and argues that the public report understated both the estimated acceleration and the uncertainty.

METR agrees with Anthropic’s assessment that acute catastrophic risk is currently low, while warning that it cannot rule out another sharp acceleration in the next six months. The nonprofit had also been asked to review how Anthropic divides compute between capability and safety research, but Anthropic ended that part of the work midway through, saying it wanted to prioritize risk assessment and revisit the compute review later. Researcher Daniel Kokotajlo argued that auditing the allocation is necessary to check whether commitments to slow development can be verified. For now, METR’s findings point to a faster research cycle—and to substantial limits on how independently the public can evaluate the evidence behind it.