🔍 Read the full analysis: AI Development At Anthropic: Findings, Questions, And Redactions on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
METR’s November 5 assessment estimates that Anthropic’s progress on measured AI capabilities accelerated about 1.9 times after October 2025, with a 95% confidence interval of 1.5x to 2.8x. METR attributes perhaps 80% of the acceleration to AI-assisted research and engineering, but says much of its most informative evidence was redacted, limiting public scrutiny.
METR, a nonprofit evaluator of frontier AI systems, published an assessment on November 5 estimating that Anthropic’s progress on measured model capabilities has accelerated to about 1.9 times its earlier pace. The analysis draws on 11 weeks of access to internal company data and says AI-assisted research likely accounts for most of the increase; METR also reports that Anthropic redacted a majority of the evidence it found most informative, leaving the public with limited ability to check key details.
METR estimates Anthropic’s capability scores rose at a rate of about 29 points a year after an October 7, 2025, break point, compared with about 15 points a year beforehand. Its estimated acceleration is 1.9x, with a 95% confidence interval of 1.5x to 2.8x. METR said it would not be surprised by estimates from 1.3x to 3x. The measure, called AECI, combines scores from multiple benchmarks into a single capability index. METR describes 2x acceleration as making two years of historical progress in one year; it is not a direct measure of researchers’ subjective productivity.
For the assessment, about 2.5 full-time-equivalent staff worked with company data, including benchmark results for unreleased models, training metrics, surveys and roughly 15 staff interviews. The work took place on Anthropic devices. METR said it received no payment; Anthropic provided laptops and free model tokens and retained the right to redact non-public material. According to the account of the report, those redactions covered a majority of the evidence METR considered most informative.
METR estimates that perhaps 80% of the acceleration is attributable to AI, while saying it is confident the change did not result from faster growth in training compute. Separately, Anthropic’s own measurement says Claude led 26% of its AI research and development work in August 2026, up from under 1% in February. The company reported about 30,000 agents working at any one time on its main internal platform. These are company figures cited in the source material, not independently verified counts in the public account.
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
AI Tools Are Becoming Part of AI R&D
The report bears on whether AI systems are helping produce more capable successors, a development that could shorten the time available for evaluation and safety work if it continues. METR’s estimate is an assessment of measured capability progress at one company, not proof that AI systems can independently design and build their successors. Anthropic’s definition of recursive self-improvement requires a system to do that work fully autonomously; the account says humans still set goals and guide development.
The findings also matter for Anthropic’s own safety commitments. Its Responsible Scaling Policy version 3.4, effective July 8, 2026, identifies automated research and development as a trigger for additional safeguards. One trigger is a doubling of progress plausibly attributable in substantial part to research or engineering automation. METR says it is plausible this threshold was crossed by July 15, though that is an interpretation, not a confirmed determination that the policy trigger has been met.
For readers, the report presents both a potentially consequential measurement and a transparency problem: the headline estimate is public, but much of the evidence behind it is not. That makes independent comparison and scrutiny harder, especially when the result may inform decisions about safeguards and the pace of AI development.
Products Worth Considering
Samsung Galaxy Tab A11 8.7-Inch Wi-Fi Tablet, 64GB, 4GB RAM, Silver
Samsung Galaxy Tab A11 Wi-Fi Tablet, 8.7-Inch, 64GB, 4GB RAM, Gray (SM-X133, International Version)
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
As an affiliate, we earn on qualifying purchases.
Anthropic’s Twofold Progress Threshold
Anthropic’s Responsible Scaling Policy sets out conditions for adding safeguards as capabilities advance. The policy describes a progress-rate threshold as making in one year the amount of progress that would otherwise take two years, provided the increase is plausibly substantially attributable to automated research or engineering. Anthropic’s Risk Report calls this a potential early warning, rather than proof that a danger is already present.
The company disclosed a related concern in an SEC filing on November 2, saying it “may imminently meet” its definition of dramatically accelerated progress. That statement preceded METR’s public assessment and is not the same as a confirmed finding that the policy threshold has been crossed.
METR’s estimate is higher than the 1.7x figure in Anthropic’s Haiku 5.5 system card, according to the source account. METR says its inclusion of newer, unreleased models is the main reason for the difference. When it instead compares frontier models across developers, the estimated acceleration is lower: 1.4x, with annual index growth changing from about 17 to 24 points. METR offers two possible explanations: Anthropic may have lagged the frontier in 2025, or its earlier models may have been less tuned to benchmarks.
““perhaps ≈80%””
— METR
Redactions Limit Public Verification
Most of the evidence METR found most informative was redacted, according to the source account. The public material therefore does not allow readers to inspect all underlying metrics, model results or interview evidence. Anthropic’s right to redact non-public material was part of the access arrangement; the reasons for each specific redaction and their effect on the estimate are not fully described in the supplied account.
The 1.9x figure is an estimate, not a direct count of research output. It changes under different assumptions: METR’s estimate ranges from 1.1x to 2.9x when assumptions about task difficulty change, while several other checks remain near 2x. Comparing Anthropic alone with the broader frontier also produces a smaller estimate of 1.4x. These differences make the choice of models, benchmarks and comparison period consequential.
It is also unclear from the available material how METR calculated the claim that AI accounts for perhaps 80% of the acceleration, or how much of the reported agent use translated into faster completed research. METR dropped a planned check of how Anthropic divides compute between safety work and capability development. The public account does not establish whether Anthropic’s policy threshold was formally triggered or what safeguards, if any, followed.
Policy Thresholds and Further Evidence
The next questions are whether Anthropic will clarify its assessment under the Responsible Scaling Policy and whether additional evidence can be released in a form that permits independent review. METR’s report offers a measured estimate, but its redactions leave room for further scrutiny of the data and methods.
Anthropic’s November 2 filing and the policy’s stated threshold make the company’s own next steps relevant. The source material does not give a confirmed date for a policy determination, a new METR assessment or further disclosure. Until those details emerge, whether Anthropic has crossed its formal threshold remains unresolved, as does the extent to which AI assistance caused the measured acceleration.
Key Questions
What did METR estimate about Anthropic’s AI progress?
METR estimated that Anthropic’s measured capability progress accelerated about 1.9x after an October 2025 break point, compared with the earlier rate. Its 95% confidence interval was 1.5x to 2.8x.
Does the report show that AI is autonomously building the next AI?
No. The report concerns AI-assisted research and development. The source account says humans still choose goals and guide the work; it does not establish fully autonomous development of a successor system.
How much of the acceleration did METR attribute to AI?
METR said perhaps 80% may be due to AI. That is an approximate estimate, not a confirmed causal measurement.
Why are the findings difficult to verify?
Anthropic redacted a majority of the evidence METR found most informative, according to the source account. The public record therefore does not expose all the information needed to independently assess the analysis.
Has Anthropic confirmed that its safety-policy threshold was crossed?
No such confirmation appears in the supplied material. METR said it was plausible the threshold had been crossed by July 15, while Anthropic’s November 2 filing said it “may imminently meet” its definition.
Source: ThorstenMeyerAI.com
