🔍 Read the full analysis: How METR Studied Whether AI Can Speed Up Its Own Development 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 capability progress on well-specified tasks accelerated about 1.9 times after October 2025. METR attributes much of the increase to AI-assisted research, but says Anthropic redacted most of the evidence it found most informative, limiting public scrutiny.
METR, a nonprofit that evaluates frontier AI systems, said on November 5 that Anthropic’s progress on well-specified AI capability tasks has accelerated to about 1.9 times its earlier rate. The estimate draws on 11 weeks of access to internal company data; METR said it believes AI use accounts for most of the acceleration, while also warning that Anthropic redacted a majority of the evidence its evaluators considered most informative.
METR’s central estimate compares Anthropic’s progress before and after a modelled break point of October 7, 2025. Using Anthropic’s capability index, the evaluators estimate that scores rose by about 15 points per year before that date and about 29 points per year afterward. METR describes the change as roughly two years of historical progress occurring in one year. Its 95% confidence interval for the acceleration is 1.5 to 2.8 times, and it said it would not be surprised by estimates between 1.3 and 3 times.
The report’s estimate is not a direct measure of how much more productive individual researchers feel. It tracks capability progress on benchmarked tasks. METR says it guesses that about 80% of the acceleration may be due to AI assistance and is confident it did not result from faster growth in training compute. That attribution is an evaluator judgment, not a directly measured division of causes.
METR staff worked with unreleased-model benchmark results, training metrics, surveys and about 15 staff interviews. Roughly 2.5 full-time-equivalent evaluators conducted the work on Anthropic devices between August 12 and October 29. METR says it received no payment; Anthropic supplied laptops and free model tokens and retained the right to redact non-public information. The report says most of the evidence METR found most informative was redacted.
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 assessment addresses whether AI systems are helping researchers develop more capable successors, a question with implications for the pace of AI progress and the time available to evaluate risks and set safeguards. If AI tools materially shorten research and engineering work, improvement could proceed faster than it would through human effort and compute growth alone. METR’s estimate offers a measurement of that possibility at one company, rather than proof that the same pattern applies across the industry.
The result also bears on Anthropic’s own safety policy. Its Responsible Scaling Policy identifies a doubling in the rate of progress, plausibly attributable to automated research or engineering, as one trigger for additional safeguards. METR says it is plausible the relevant threshold was crossed by July 15, 2026. That is METR’s assessment, not a definitive finding that the policy trigger has been met.
For readers and policymakers, the report illustrates both the value and limits of outside evaluation using company records. Access to internal data can reveal trends that public model releases cannot show. But redactions make it harder for independent researchers to reproduce the analysis or judge how strongly the evidence supports the estimate.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, Touch ID, Wi-Fi 7, Midnight
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Anthropic’s Twofold Progress Threshold
Anthropic’s Responsible Scaling Policy version 3.4 took effect on July 8, 2026. It identifies two possible signs of automated research and development: an AI system able to replace the company’s research staff at competitive cost, or a doubling of progress within a year compared with the historical pace over two years, where the acceleration is plausibly substantially attributable to automated research or engineering.
The policy presents the doubling threshold as an early warning, not proof that a particular danger has arrived. Anthropic has said it is not yet at the point of an AI system autonomously designing and developing its own successor, and that recursive self-improvement is not inevitable. In a November 2 SEC filing, the company said it “may imminently meet” its definition of dramatically accelerated progress.
METR’s headline estimate is higher than the 1.7-times figure in Anthropic’s Haiku 5.5 system card. METR says including newer, unreleased models is a key reason for the difference. When it instead examines frontier models across developers, it estimates a smaller rise, from about 17 to 24 index points per year, or 1.4 times. METR says the difference could reflect Anthropic catching up from a weaker position or changes in how well earlier models performed on benchmarks.
““We would not be surprised” by an acceleration estimate from 1.3x to 3x.”
— METR
Evidence Behind the Estimate
The public report does not expose much of the material METR considered most informative. Anthropic redacted a majority of that evidence, and the company also dropped a planned check of how it divides compute between safety work and capability development. The available account therefore does not let outside readers fully assess the underlying records or independently reproduce the analysis.
The estimate also changes with methodological assumptions. METR says variations in assumptions about task difficulty produce results ranging from 1.1 to 2.9 times, although most of its other checks remain near twofold. The comparison across all developers yields a lower estimate than the Anthropic-only measure. The report does not establish whether Anthropic’s trend reflects a broader change across AI labs, nor does it show that AI systems are independently setting research goals or building successors without human direction.
METR’s estimate that AI accounts for about 80% of the acceleration is explicitly tentative. The source material does not provide a publicly checkable breakdown that settles how much of the change comes from AI tools, research practices, model selection or other factors. The precise status of Anthropic’s policy threshold also remains unresolved in the material presented.
Policy Threshold and Further Scrutiny
The next key development is whether Anthropic determines that its Responsible Scaling Policy’s automated-R&D threshold has been reached and what safeguards would follow. The company’s November 2 filing said it may imminently meet its definition; METR’s report says a crossing by July 15 is plausible. Neither statement, on its own, settles the policy status.
Further scrutiny will depend on what evidence Anthropic makes available and whether evaluators can test the findings with fewer restrictions. The report also leaves open whether future assessments will examine compute allocation between safety and capabilities, a check that was planned but dropped for this study. Until additional information is released, the 1.9-times estimate remains an independent assessment with substantial limits on public verification.
Key Questions
What did METR estimate?
METR estimated that Anthropic’s progress on well-specified AI capability tasks accelerated by about 1.9 times after a modelled October 2025 break point, compared with its earlier pace.
Does the report say AI is developing its own successor?
No. The report concerns AI helping people with AI research and development. Anthropic says it has not reached the point where an AI system autonomously designs and develops its successor.
How much of the acceleration does METR attribute to AI?
METR tentatively estimates that perhaps about 80% is due to AI assistance. It presents this as a judgment about likely causes, not a precisely measured share.
Why is the estimate difficult to verify?
Anthropic redacted a majority of the evidence METR found most informative. The report also describes sensitivity to assumptions, with some task-difficulty models producing estimates from 1.1 to 2.9 times.
Has Anthropic confirmed that its policy threshold was crossed?
The available source material does not confirm that it was. METR says a crossing by July 15 is plausible, while Anthropic said in a November 2 filing that it may imminently meet its definition of dramatically accelerated progress.
Source: ThorstenMeyerAI.com
