🔍 Read the full analysis: What Did METR Find Inside Anthropic 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 measured AI capabilities accelerated about 1.9x after October 2025, with a 95% confidence interval of 1.5x to 2.8x. The nonprofit says AI use likely accounts for most of the change, but Anthropic redacted much of the evidence and METR did not complete a planned analysis of compute devoted to safety versus capabilities.
METR, an AI evaluation nonprofit, estimates that Anthropic’s measured model capability progress accelerated to about 1.9 times its earlier pace after October 2025, according to an assessment published November 5. The finding matters because METR believes AI tools likely account for most of the acceleration—but the company redacted a majority of the evidence the evaluators considered most informative, limiting what outsiders can verify.
METR staff spent 11 weeks examining internal Anthropic data, including benchmark results for unreleased models, training metrics, surveys and about 15 staff interviews. The work ran from August 12 to October 29, 2026, with about 2.5 full-time-equivalent staff working on Anthropic devices. METR says it received no payment; Anthropic supplied laptops and free model tokens and retained the right to redact non-public material.
METR’s main estimate compares the pace of gains on a capability index before and after a breakpoint it places at October 7, 2025. It estimates the index rose about 15 points per year before that date and about 29 points per year afterward. The resulting 1.9x estimate has a 95% confidence interval of 1.5x to 2.8x. METR says it would not be surprised by a result between 1.3x and 3x, and reports that alternative assumptions about task difficulty produce estimates from 1.1x to 2.9x.
The report attributes perhaps about 80% of the acceleration to AI use, while saying it is confident the change did not result from faster growth in training compute. That is METR’s interpretation, not a direct measurement of the share of work performed by AI. Anthropic’s own internal measure, cited in the report, says Claude led 26% of its AI research and development work in August 2026, compared with less than 1% in February; about 30,000 agents were working at a time on its main internal platform.
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 Faster AI Research Matters
The result bears on whether AI systems are beginning to speed up the development of later systems. If AI tools help researchers build, test or improve models more quickly, advances could arrive faster than they would through human work and additional computing resources alone. METR’s estimate suggests this process may already be measurable at Anthropic, though it does not establish that AI systems are autonomously creating their successors.
The finding also has direct relevance to Anthropic’s own safety policy. The company’s Responsible Scaling Policy includes a threshold for a doubling in the rate of progress when that acceleration is plausibly tied substantially to automated research or engineering. METR says it is plausible that Anthropic crossed the policy’s early-warning threshold by July 15. Anthropic stated in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. These are related assessments, but neither statement alone establishes that the policy threshold has definitively been met.
For readers and policymakers, the report’s limitations are part of the news. A detailed external review of a lab’s internal development process is unusual, but the redactions restrict independent scrutiny. The public can see METR’s conclusion and some of its methods, but not much of the evidence it says was most useful in reaching that conclusion.
Popular right now
Breville Bambino Espresso Machine, Brushed Stainless Steel
De’Longhi Classic Matte Black Espresso Machine with Milk Frother
Hisense 43-Inch E6SR Series Hi-QLED 4K UHD Smart Roku TV (2026 Model)
As an affiliate, we earn on qualifying purchases.
METR’s Internal Data Review
METR’s assessment differs from the nonprofit’s usual evaluations of finished AI models. In this project, evaluators examined company-held information, including scores from unreleased Anthropic models and internal measures of training and research. The report says the team also surveyed employees and interviewed about 15 key staff members. Its estimate is based on the observed rate of capability gains, not a direct count of all research tasks completed by people or AI agents.
Anthropic’s Responsible Scaling Policy, version 3.4, took effect July 8, 2026. It defines two possible triggers for automated research and development: a model capable of replacing the company’s research staff at competitive cost, or a doubling in the pace of progress that is plausibly substantially attributable to automated research or engineering. The policy describes this as an early warning, not proof that the most extreme form of self-improving AI exists.
Anthropic has said it is not yet at fully autonomous recursive self-improvement, which it describes as an AI system independently designing and developing its successor. The situation assessed by METR is more limited: AI tools assist human researchers, who continue to set goals and direct the work.
“Perhaps ≈80% of the current acceleration is due to AI.”
— METR
What Redactions Leave Unknown
METR says Anthropic redacted a majority of the evidence it found most informative. The report therefore does not provide the public with a full basis for independently checking its interpretation. METR also dropped a planned examination of how Anthropic divides computing resources between safety work and capability development, leaving that question unresolved.
The 1.9x figure is an estimate based on a particular capability index, selected models and assumptions about task difficulty. METR reports a smaller acceleration—1.4x, from 17 to 24 index points per year—when comparing frontier models across developers rather than Anthropic models alone. It offers two possible explanations: 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 benchmarks, which could make the measured change look smaller. The available material does not settle between these explanations.
It is also not clear from the cited findings how much the reported shift translates into shorter end-to-end development timelines, or whether it will continue. An acceleration in benchmark capability progress is not the same as researchers feeling twice as productive, nor does it show that an AI system can independently choose and execute a full model-development plan.
Policy Decisions and Further Evidence
The immediate next question is how Anthropic applies its Responsible Scaling Policy to its own progress. The company’s filing said it may imminently meet the relevant definition; METR judged that crossing the early-warning threshold by July 15 was plausible. The available source material does not report a definitive policy determination following the assessment.
Further public detail would help readers test the estimate, particularly information about the redacted evidence, the models included in the capability trend and the analysis METR could not complete on safety versus capability compute. Any later update should also clarify whether Anthropic’s measured pace continues, slows or changes under different evaluation methods. Until then, the report is a significant but qualified measurement: a near-doubling estimate paired with substantial limits on public verification.
Key Questions
What did METR estimate about Anthropic’s AI progress?
METR estimated that Anthropic’s measured capability progress accelerated 1.9x after a breakpoint it places at October 7, 2025. It gives a 95% confidence interval of 1.5x to 2.8x.
Does the report show that AI is developing AI on its own?
No. METR describes AI as helping people carry out research and development. Anthropic says it has not reached fully autonomous recursive self-improvement, in which an AI system independently designs and develops its successor.
How much of the acceleration does METR attribute to AI?
METR says perhaps about 80% of the acceleration is due to AI. That is the nonprofit’s estimate, not a directly measured proportion of all research work.
Why is the finding difficult to verify independently?
METR says Anthropic redacted a majority of the evidence it found most informative. It also did not complete a planned analysis of how the company divides computing resources between safety and capability work.
Has Anthropic confirmed it crossed its policy threshold?
The source material does not establish a definitive crossing. Anthropic said in a November 2 SEC filing that it may imminently meet its definition, while METR said crossing the early-warning threshold by July 15 was plausible.
Source: ThorstenMeyerAI.com
