TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
METR’s review of Anthropic data estimates that the company’s model test scores have improved about 1.9 times faster since October 2025 than in the preceding period. The nonprofit says roughly 80% of the acceleration may stem from AI use in Anthropic’s research, and estimates the company may have crossed its own warning threshold in July 2026. The findings are uncertain, and extensive redactions prevent readers from independently checking key evidence.
Nonprofit METR says Anthropic may have crossed its own warning threshold for AI-driven acceleration as early as July 15, 2026, after the company’s model test scores began rising about 1.9 times faster than before. METR estimates that roughly 80% of the speedup is linked to AI being used in Anthropic’s research, but says extensive company redactions leave the public unable to inspect much of the evidence behind the conclusion.
METR based its review on eleven weeks of access to Anthropic’s internal research data, from August 12 through October 29, 2026. The nonprofit says it examined benchmark results for released and unreleased models, training metrics, employee surveys and internal analyses, and interviewed about 15 Anthropic researchers. Three unreleased models included in the analysis completed training between July and October.
Using the AECI composite benchmark, METR calculated that Anthropic’s models gained about 15 points per year before October 2025 and about 29 points per year afterward. It describes the recent pace as roughly 1.9 times faster than the rate from March 2024 to October 2025. METR says the change is not explained by faster growth in training compute and estimates that AI-assisted research accounts for most of it.
The review does not find that AI systems are conducting Anthropic’s research autonomously. METR says the gains mainly come from models assisting researchers, while describing some signs that independent contributions could grow. Its estimate is imprecise: the acceleration factor could be between 1.3 and 3, according to the nonprofit. When it includes models from other companies, METR’s estimate of industry-wide acceleration falls to about 1.4.
Why AI-Assisted Research Matters
The finding matters because faster AI development could change how quickly developers need to evaluate model capabilities and risks. Anthropic’s Responsible Scaling Policy treats a doubling in the pace of progress attributable to automated AI research as a warning indicator. METR says its estimate is near that level and that the threshold may have been reached in July.
That is a warning about a possible trend, not evidence that a runaway feedback loop is already underway. METR says the current acceleration is mainly from AI acting as a research assistant, not from systems independently designing and building successor models. Still, the nonprofit found no clear bottlenecks that it expects to slow progress over the next 12 months, and said there is a significant risk the acceleration could continue to grow.
The estimates also bear on public oversight. If AI tools materially speed up the work that produces more capable systems, the pace of capability gains could become harder to forecast using earlier trends. But the uncertainty and the limited public evidence mean the report should not be read as a precise measurement of how quickly the entire field is advancing.
Products Worth Considering
Samsung Galaxy Tab A11 Wi-Fi Tablet, 8.7-Inch, 64GB, 4GB RAM, Gray (SM-X133, International Version)
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
Anthropic’s Policy and METR Review
Anthropic’s August 2026 risk report said the acceleration was less than a factor of two. METR says the company’s unredacted internal version implies a figure of 1.97 and that Anthropic itself described crossing the threshold as plausible. METR estimates the threshold may have been reached on July 15 and argues that the public report understated both the estimated acceleration and its uncertainty.
METR is a nonprofit that studies risks associated with advanced AI, including whether developers could lose control of their systems. It says it received no money from Anthropic for this review, although its team worked on Anthropic devices for security reasons. The review is separate from an October collaboration between the organizations, which Anthropic described as involving “employee-like access”; METR says that separate work has a different and narrower scope.
Anthropic had unilateral authority to redact non-public material from the review. METR says the company removed much of the evidence the auditors considered most informative, including future-model forecasts, training-step breakdowns and bottleneck analysis. Anthropic said the redactions were “aimed at removing confidential information that could accelerate other AI developers.” METR says this leaves readers relying on its interpretation more than it considers ideal.
““Aimed at removing confidential information that could accelerate other AI developers.””
— Anthropic, describing its reason for redacting material, as quoted by METR
Redactions Limit Independent Review
The size and cause of the acceleration remain uncertain. METR places the possible rate between 1.3 and 3 times the earlier pace, and researchers interviewed by the nonprofit disagreed about how much AI contributed. One researcher estimated AI’s share of productivity gains at only 15% to 20%; another identified new hires as the leading factor in recent progress.
METR also cautioned against treating employee surveys as firm evidence. Informal polls in October 2026 suggested employees felt roughly 12 times more productive with an internal model than with Claude 3.7 Sonnet, but the nonprofit called these self-assessments unrepresentative and likely to overstate actual gains. The result is not a controlled measure of productivity across the company.
Most importantly, readers cannot check key parts of the audit because Anthropic removed material METR rated as highly informative. METR says the redactions include forecasts, training breakdowns and its analysis of bottlenecks. It is also unclear from the available material how much the reported acceleration reflects a catch-up effect: METR says Anthropic may have been behind competitors in 2025, which could skew comparisons.
Further Audit Findings to Watch
The report identifies a possible warning-level acceleration, but the available material does not establish that Anthropic has formally activated a policy response or changed its development plans. The next points to watch are whether Anthropic publishes further explanation of its threshold assessment, whether METR provides additional evidence that can be reviewed publicly, and how the company’s rate of model progress changes over the coming year.
METR also reviewed Anthropic’s August 2026 risk report in a companion assessment. The nonprofit’s current view is that acute catastrophic risk has not been established by this finding, while the possibility of continued acceleration deserves attention. Future model evaluations and clearer disclosures would help show whether the trend persists and how much AI assistance contributes.
Key Questions
What did METR find about Anthropic’s AI progress?
METR estimates that Anthropic’s model scores on a composite benchmark rose about 1.9 times faster after October 2025 than during the preceding period. It attributes roughly 80% of the estimated acceleration to AI use in the company’s research, while stressing that the estimate is uncertain.
What warning threshold may Anthropic have crossed?
Anthropic’s Responsible Scaling Policy treats a doubling in the pace of progress because of automated AI research as a warning indicator. METR says the threshold may have been reached around July 15, 2026; it does not say the timing or estimate is certain.
Are Anthropic’s AI systems doing research autonomously?
METR says the current acceleration mainly comes from AI assisting researchers, not from fully autonomous AI research. The nonprofit says it found signs that independent contributions could become more significant, but does not report that this has already happened at scale.
How reliable is the estimate?
METR says the acceleration factor could range from 1.3 to 3. It also says Anthropic redacted much of the evidence auditors considered most informative, limiting public scrutiny. Researchers interviewed by METR differed on AI’s role, and employee productivity responses were informal self-reports.
Does the report show that an AI safety emergency has occurred?
No. The report identifies a possible acceleration warning signal and a risk that the trend could continue. It does not establish that AI systems are autonomously developing successors or that a catastrophic event is underway.
Source: the-decoder.com
