🔍 Read the full analysis: METR’s Anthropic Findings Raise Questions About AI Development on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
METR’s November 5 assessment estimates that Anthropic’s progress on well-specified AI tasks accelerated about 1.9 times from its pace between March 2024 and October 2025. METR attributes perhaps 80% of the change to AI assistance, but says Anthropic redacted most of the evidence it found most informative, leaving important details difficult for outsiders to verify.
METR, a nonprofit that evaluates frontier AI systems, estimates that Anthropic’s progress on measured AI capabilities has accelerated about 1.9 times compared with its pace from March 2024 to October 2025. The November 5 assessment draws on 11 weeks of access to internal company data and suggests that AI tools account for much of the faster progress, but METR says Anthropic redacted most of the evidence it found most informative, limiting what outsiders can independently assess.
METR measured progress using Anthropic’s version of the Epoch Capabilities Index, a composite of benchmark results. It estimates the score rose by about 15 points a year before October 2025 and about 29 points a year afterward. METR places the change point at October 7, 2025, and describes the result as roughly two years of earlier progress occurring in one year. Its estimated acceleration is 1.9 times, with a 95% confidence interval of 1.5 to 2.8 times.
The report’s estimate is not a direct measurement of how much faster every part of AI development has become. METR says it is measuring capability progress on well-specified tasks, not overall research productivity or the time required to complete the whole development process. It estimates that perhaps about 80% of the acceleration is attributable to AI use, while saying it is confident the change did not result from faster growth in training compute. Those are METR’s judgments, not independently established facts about the cause.
METR reports that its review involved about 2.5 full-time-equivalent staff working on Anthropic devices, examining benchmark scores for unreleased models, training metrics, surveys and about 15 staff interviews. The organization says it received no payment; Anthropic provided laptops and free tokens and retained the right to redact non-public material. METR also says it dropped a planned examination of how Anthropic divides compute between safety work and capability development.
Infographic 01 · How fast is AI speeding up AI?
About twice the old pace, most likely because of AI
An outside evaluator's estimate from internal Anthropic data, plus Anthropic's own measurements. Wide uncertainty, much of the evidence redacted.
Acceleration of Anthropic’s benchmark progress (AECI) since October 2025, 95% confidence interval 1.5x to 2.8x. METR’s overall judgment: around twice the old rate; it “would not be surprised” by 1.3x or 3x.
Capability points per year
AECI trend of Anthropic models, break point 7 Oct 2025
All labs before October 2025: +17 a year, so 1.4x. Different task-difficulty assumptions: 1.1x to 2.9x.
Sources: METR assessment, 5 Nov 2026 · METR risk report review · Anthropic Institute measurements.
Powered by Thorsten Meyer AI
What Faster AI Progress Could Mean
The report addresses whether AI systems are helping researchers build more capable successor systems faster. That question matters to safety planning because a faster development cycle could shorten the time available to evaluate models, respond to new capabilities and put safeguards in place. METR’s estimate offers an outside assessment of one company’s internal development process rather than relying only on public model releases.
The findings do not establish that AI can independently design and build its own successor. The source material describes current work as AI assisting human researchers, with people still choosing goals. Anthropic has described full autonomous development of a successor as a form of recursive self-improvement and said it is not there yet. The distinction matters: evidence that AI contributes to research is not proof that the entire process has become autonomous.
METR’s estimate also bears on Anthropic’s own safety framework. The company’s Responsible Scaling Policy identifies a doubling in progress, plausibly attributable to automated research or engineering, as one possible trigger for added safeguards. METR says it is plausible that the threshold may have been crossed by July 15, but that conclusion depends on estimates and definitions. It is not a finding that the policy trigger was definitively met.
Popular right now
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green
Hisense 40-Inch Class QD4SR Series Hi-QLED FHD Smart Roku TV (2026 Model)
As an affiliate, we earn on qualifying purchases.
How METR Measured Anthropic’s Progress
METR’s work differs from an evaluation that tests only a publicly released model. The nonprofit examined internal material, including results from unreleased systems, and worked with Anthropic staff during the review. Its central comparison uses a trend before and after an estimated October 2025 break point. The result therefore depends on the benchmarks included, the way they are combined and the choice of baseline period.
METR reports a smaller acceleration when it compares frontier models across developers rather than Anthropic models alone: about 1.4 times, based on a change from roughly 17 to 24 index points per year. It offers two possible interpretations: Anthropic may have been behind the frontier in 2025, making its later catch-up look larger, or its earlier models may have been less tuned to the benchmarks. Under different assumptions about task difficulty, the estimate ranges from 1.1 to 2.9 times; METR says most of its other checks remain near two times.
The supplied source also reports Anthropic’s own internal measure that Claude led 26% of its AI research and development work in August 2026, compared with less than 1% in February. It says about 30,000 agents were operating at a time on the company’s main internal platform. These are figures attributed to Anthropic, and “led” is not defined here as meaning that Claude performed all of the work or acted without human direction.
““Perhaps ≈80%” of the current acceleration is due to AI.”
— METR
Redactions Limit Independent Checks
The largest limitation is access to the underlying evidence. METR says Anthropic redacted a majority of the evidence it found most informative, and the company retained the right to withhold non-public material. The assessment is independent in the sense that METR conducted it and says it took no payment, but readers cannot review all the material that informed its conclusions.
Several details remain unresolved from the source material: how the redacted evidence might affect the estimate, how much of the measured acceleration reflects changes in benchmark performance rather than broader research capability, and how Anthropic divides resources between safety and capabilities. METR did not complete its planned compute-allocation check. The reported share of progress attributable to AI is also an estimate, not a demonstrated causal breakdown.
There is a further difference between the company-wide frontier comparison and Anthropic-only estimate. METR’s 1.9-times figure becomes about 1.4 times when it looks across developers. The report offers explanations for that gap but, on the information provided, does not settle which explanation is correct.
Further Evidence Will Shape the Threshold
The source material does not identify a scheduled follow-up assessment or a date when additional evidence will be released. What readers can watch for is whether METR publishes further details or robustness checks, whether Anthropic makes more of the underlying evidence available, and whether the company updates its assessment under its Responsible Scaling Policy.
Any later claim that Anthropic has crossed its policy threshold would need to be distinguished from METR’s current finding that crossing it by July 15 is plausible. Until the evidence and methods can be examined more fully, the 1.9-times estimate is a significant outside assessment, but its exact meaning and causes remain open to scrutiny.
Key Questions
What did METR conclude about Anthropic’s AI progress?
METR estimated that progress on its measured capability index accelerated 1.9 times after an October 2025 change point, relative to the earlier trend. It reported a 95% confidence interval of 1.5 to 2.8 times.
Does the report show that AI is building AI autonomously?
No. The source describes AI assisting human researchers and says humans still choose development goals. It does not establish that a system can independently design and build its own successor.
How much of the acceleration does METR attribute to AI?
METR estimates that perhaps about 80% is due to AI use. That is the organization’s attribution, not a directly verified causal measurement.
Why are the findings difficult to verify?
METR says Anthropic redacted a majority of the evidence it considered most informative. The company also retained the right to redact non-public material, and METR did not complete a planned review of compute allocation between safety and capability work.
Did METR say Anthropic had definitely crossed its safety-policy threshold?
No. METR said it was plausible the threshold may have been crossed by July 15. That is not a definitive finding that Anthropic’s policy trigger was met.
Source: ThorstenMeyerAI.com
