🔍 Read the full analysis: AI And Its Own Development: What METR Saw At Anthropic on ThorstenMeyerAI.com
Get hardware and tech essentials 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 estimates that Anthropic’s capability progress on well-specified tasks accelerated to about 1.9 times its earlier pace, with AI assistance likely responsible for most of the increase. The assessment draws on 11 weeks of internal data, but Anthropic redacted a majority of the evidence METR found most informative, limiting independent public scrutiny.
METR, an AI evaluation nonprofit, estimates that Anthropic has been improving models on well-specified tasks at about 1.9 times its earlier pace, in an assessment published November 5. Anthropic’s public offering filing also discusses risks related to AI. The report says AI systems likely account for most of that acceleration, but its public findings are difficult to independently verify because Anthropic redacted a majority of the evidence METR considered most informative.
METR’s main estimate compares Anthropic’s capability progress before and after a change point it places at October 7, 2025. Its measure, AECI, combines scores across benchmarks. The report estimates the score rose by about 15 points per year before the change point and about 29 points per year afterward. METR describes that 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.
The report’s broader judgment is less exact: METR says it would not be surprised by an acceleration between 1.3 and 3 times. For another example of Anthropic’s evolving AI systems, see Claude’s team of agents. It estimates that “perhaps ≈80%” of the current acceleration is due to AI assistance, and says it is confident the change did not come from faster growth in training compute. That attribution is METR’s assessment, not a directly observable breakdown of every cause.
METR says its team spent 11 weeks working with internal Anthropic data, including benchmark scores for unreleased models, training metrics, surveys and roughly 15 staff interviews. About 2.5 full-time-equivalent staff worked on the assessment from August 12 to October 29 using Anthropic devices. METR says it received no payment; Anthropic provided laptops and free tokens and retained the right to redact non-public information. METR also says Anthropic did not provide the planned evidence on how compute is divided between safety and capability work.
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 report bears on whether AI is beginning to speed up the research and engineering that produce more capable AI systems. If that contribution grows, development could move faster than it would through human work and computing resources alone. That possibility matters to safety planning because evaluation, safeguards and governance may need to keep pace with the systems they are intended to address.
METR’s finding is not that Anthropic’s systems can independently design and build their own successors. Anthropic’s description of recursive self-improvement refers to a system that can fully autonomously develop its successor, and the company says current systems have not reached that point. The report instead concerns AI helping people conduct research and engineering. Its estimate therefore informs an early-warning discussion, rather than establishing that fully autonomous self-improvement has begun.
The evidence limits matter alongside the estimate. A result based partly on unreleased models and internal company data can provide information unavailable in public benchmark testing, but redactions make it harder for outside researchers to assess the methods, assumptions and supporting evidence. The reported acceleration is a quantified assessment, not a complete public audit.
Products Worth Considering
Apple MacBook Pro 16-inch with M1 Pro, 10-core CPU, 16-core GPU, 16GB RAM, 1TB SSD, Silver
Apple MacBook Pro 16.2-inch with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 1TB SSD, Nano-Texture Display, Space Black
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
Anthropic’s Two-Times Progress Trigger
Anthropic’s Responsible Scaling Policy, version 3.4, took effect on July 8, 2026. It defines triggers for “automated R&D,” including either a model capable of replacing the company’s research staff at competitive cost or a doubling of the pace of progress plausibly attributable in substantial part to automated research or engineering. Under the policy’s definition, a 2x acceleration means achieving in one year the amount of progress that would otherwise take two years.
Anthropic’s Risk Report describes the 2x level as a potential early warning, not proof that a particular danger is already present. METR says it is plausible the threshold had 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 offers an outside estimate, while the filing describes Anthropic’s own assessment under its policy.
METR’s result also depends on what is measured. When it assessed frontier models across developers rather than Anthropic models alone, the estimated pace rose from about 17 to 24 AECI points per year, or 1.4 times. METR says Anthropic may have been behind the frontier in 2025, which could make its company-specific acceleration look larger; alternatively, earlier models may have been less tuned to benchmarks, which could make the estimate smaller. Under different task-difficulty assumptions, the estimate ranged from 1.1 to 2.9 times.
““perhaps ≈80%””
— METR
Redactions Limit Public Verification
The main unresolved issue is how much of METR’s evidence can be checked publicly. The report says Anthropic redacted a majority of the material METR found most informative. The source material does not specify exactly which findings or supporting records were removed, so readers cannot fully judge how those omissions affect the analysis.
The 1.9x figure is an estimate based on a particular benchmark composite, a selected change point and company data that includes unreleased models. METR reports meaningful sensitivity to assumptions about task difficulty and presents a lower estimate when examining frontier models across developers. The assessment does not establish a single cause for the measured change or prove that AI assistance accounts for exactly 80% of it.
It is also unclear from the published material how Anthropic will apply its policy threshold in practice, what safeguards would follow if it determines a trigger has been met, or how its work is split between capability development and safety. METR’s planned check of that compute allocation did not go ahead.
Policy Decisions and Further Evidence
The next key development is whether Anthropic formally determines that it has met a trigger in its Responsible Scaling Policy and, if so, what steps it takes under that policy. Its November 2 filing said the threshold may be imminent; the material cited here does not report a final determination or a specific timeline for one.
Further public evidence could clarify the estimate’s assumptions and its implications, particularly if more of the redacted material becomes available or later assessments test the same trends. Until then, METR’s report offers an unusually detailed outside estimate based on internal access, but the extent of the acceleration and AI’s precise contribution remain uncertain.
Key Questions
What did METR estimate about Anthropic’s AI progress?
METR estimated that Anthropic’s capability progress on well-specified tasks accelerated by about 1.9 times compared with its earlier pace. Its reported 95% confidence interval is 1.5 to 2.8 times.
Does the report say AI is building AI without human involvement?
No. The report concerns AI systems helping people with AI research and engineering. It does not establish that a system can autonomously design and develop its own successor; Anthropic says that stage has not been reached.
How much of the acceleration does METR attribute to AI?
METR estimates that “perhaps ≈80%” comes from AI assistance. That is a qualified estimate, not a definitive accounting of every cause.
Why can’t readers fully verify the findings?
METR says Anthropic redacted a majority of the evidence it found most informative. The report also relies partly on internal data about unreleased models, and the published material does not identify all the redacted evidence.
Has Anthropic confirmed that its policy threshold was crossed?
The supplied material does not establish a final determination. METR says it is plausible the threshold was crossed by July 15, while Anthropic said in a November 2 filing that it “may imminently meet” its definition.
Source: ThorstenMeyerAI.com
