🔍 Read the full analysis: What METR’s Anthropic Research Says About AI And Its Own Progress on ThorstenMeyerAI.com
Turn your quiet moments into listening time
- Thousands of audiobooks, podcasts and originals
- Listen on your phone, tablet or Echo — also offline
- Cancel anytime
TL;DR
METR estimates that Anthropic’s capability progress on well-specified tasks accelerated about 1.9 times after October 2025, with AI assistance likely accounting for most of the change. Its assessment drew on internal data, but Anthropic redacted much of the evidence METR considered most informative, leaving limits on what outsiders can verify.
METR, a nonprofit that evaluates frontier AI systems, estimates that Anthropic’s progress on measured AI capabilities accelerated about 1.9 times after October 2025, compared with its pace from March 2024 through October 2025. The independent assessment, published November 5 and based on access to internal company data, says AI assistance likely accounts for most of the acceleration—but Anthropic redacted a majority of the evidence METR found most informative, limiting public scrutiny.
METR’s central estimate is that Anthropic’s measured progress rose from about 15 points per year to 29 points per year on the company’s capability index, a 1.9x change. METR identifies October 7, 2025, as the estimated break point. It describes the result as roughly two years of historical progress occurring over one year. The report gives a 95% confidence interval of 1.5x to 2.8x and says METR would not be surprised by an estimate between 1.3x and 3x.
The measure, AECI, combines scores from multiple benchmarks into a single capability index. METR says its estimate is higher than the 1.7x reported in Anthropic’s Haiku 5.5 system card, mainly because METR included newer, unreleased models. Looking across frontier models from all developers, rather than Anthropic’s models alone, METR estimates a smaller acceleration: from 17 to 24 points per year, or 1.4x. The report says the difference could reflect Anthropic catching up from behind in 2025 or changes in how its earlier models performed on benchmarks.
METR staff spent 11 weeks working with Anthropic data, including benchmark scores for unreleased models, training metrics, staff surveys and about 15 interviews. METR says about 2.5 full-time-equivalent staff worked on the assessment from August 12 to October 29, using Anthropic devices. The nonprofit says it received no payment; Anthropic provided laptops and free tokens and retained the right to redact non-public information.
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
Why AI-Assisted Research Matters
The report addresses a consequential question for AI safety: whether AI tools are helping labs develop more capable AI systems faster. If that contribution grows, the time between capability gains may shrink, affecting how quickly researchers, companies and regulators can evaluate risks and put safeguards in place. METR’s estimate is evidence about one company, not proof that the entire industry is accelerating at the same rate.
METR estimates that perhaps about 80% of Anthropic’s acceleration came from AI assistance, while saying it is confident the change did not result from faster growth in training compute. That breakdown is an estimate, not a directly established causal finding. The report also cites Anthropic’s own measurement that Claude led 26% of its AI research and development work in August 2026, up from less than 1% in February. Anthropic reportedly had about 30,000 agents working at any one time on its main internal platform. Those figures describe the company’s internal measures; they do not, on their own, establish how much the agents increased research output.
The distinction matters because faster progress on benchmarked tasks is not the same as fully autonomous AI development. Humans still set goals and make decisions in the process described by the report. METR’s estimate concerns the pace of measured capability progress, not a finding that an AI system can independently design and build its successor.
Products Worth Considering
Oura Ring 5 Smart Ring, Gold, Size 8
Garmin Forerunner 935 GPS Running and Triathlon Watch
Withings Sleep Under-Mattress Sleep Tracking Pad
As an affiliate, we earn on qualifying purchases.
Anthropic’s Twofold Progress Trigger
Anthropic’s Responsible Scaling Policy, version 3.4, took effect July 8, 2026. It defines two possible triggers for what it calls automated research and development: a model able to replace the company’s research staff at competitive cost, or a doubling of progress plausibly attributable in substantial part to automated research or engineering. The policy defines doubling as achieving in one year the progress that would previously have taken two.
Anthropic describes the threshold as a potential early warning, not proof that a danger has arrived. METR says it is plausible the threshold may have been crossed by July 15, 2026. In a November 2 SEC filing, Anthropic said it “may imminently meet” its definition of dramatically accelerated progress. These are related but distinct statements: METR’s assessment is an outside estimate, while the filing reflects Anthropic’s own disclosure.
METR’s work differs from a typical outside evaluation, which tests a released model. Here, evaluators examined internal material, including unreleased-model results and company data. That access gives the assessment a view unavailable from public benchmarks alone, while the redactions make it harder for readers to independently check the underlying evidence.
““roughly two years of historical progress over the last year””
— METR
What the Redactions Leave Unknown
The public cannot fully inspect the evidence behind METR’s conclusions. The report says Anthropic redacted a majority of the evidence METR found most informative, and it dropped a planned check of how the company divides compute between safety work and capability development. The source material does not specify all the withheld details, so readers cannot determine from the published account how each redaction affects confidence in the findings.
The 1.9x figure also depends on measurement choices. METR says alternative assumptions about task difficulty move its estimate between 1.1x and 2.9x, although most other checks remain near 2x. Its all-developer estimate is lower, at 1.4x. The report does not establish that the same acceleration applies across other AI companies, or that AI assistance alone caused the change.
It also remains unclear how much the reported increase in Claude’s share of internal work translated into faster model development. The measures described cover benchmark performance and internal research activity; they do not show that AI systems now independently control the research process.
How the Estimate Will Be Tested
The next evidence will come from whether later model evaluations and disclosures support METR’s estimated pace of progress. Anthropic’s Responsible Scaling Policy provides a stated threshold for adding safeguards, while the company’s November 2 filing indicates that it believes it may approach its own definition of dramatically accelerated progress. The public report does not specify a date for a policy decision or a further assessment.
Further scrutiny would depend on clearer disclosure of the data and methods behind the estimate, including the evidence Anthropic redacted and the omitted analysis of compute allocation. Until then, METR’s findings offer an unusually detailed outside assessment of one lab, but their precision and broader relevance remain limited by the information available for public review.
Key Questions
What did METR conclude about Anthropic’s progress?
METR estimated that Anthropic’s measured capability progress accelerated 1.9x after October 2025 compared with its earlier pace. Its reported 95% confidence interval is 1.5x to 2.8x.
Does the report say AI is building AI autonomously?
No. The report describes AI assisting human research and development. It does not establish that an AI system can independently set goals and develop its own successor.
How much of the acceleration did METR attribute to AI?
METR estimated that perhaps about 80% came from AI assistance. That is a qualified estimate, not a direct causal measurement.
Why is the result difficult to verify?
METR says Anthropic redacted a majority of the evidence it considered most informative. The report also omitted a planned check of how the company divides compute between safety and capability work.
Does the estimate apply to the whole AI industry?
Not directly. METR’s main estimate concerns Anthropic. Its comparison across frontier models from all developers found a smaller acceleration of 1.4x, and the report does not establish a shared industry-wide rate.
Source: ThorstenMeyerAI.com
