🔍 Read the full analysis: The AI Development Question: What METR Found At Anthropic on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s November 5 assessment estimates that Anthropic’s AI capabilities have improved about 1.9 times faster since October 2025 than in the preceding period. The nonprofit attributes perhaps 80% of the acceleration to AI-assisted development, but says most of its most informative evidence was redacted, leaving the public with limited ability to check the findings.
METR, a nonprofit that evaluates frontier AI systems, said on November 5 that Anthropic’s model capabilities have been improving at about 1.9 times their previous pace since October 2025. The assessment, based on 11 weeks of access to company data, estimates that AI tools likely account for much of the acceleration—but Anthropic redacted a majority of the evidence METR considered most informative, limiting what outsiders can verify.
METR compared Anthropic’s capability progress before and after a break point it estimates occurred on October 7, 2025. Using a combined benchmark measure, it estimates progress rose from roughly 15 points per year to about 29 points per year. METR’s central estimate is a 1.9x acceleration, with a 95% confidence interval of 1.5x to 2.8x. The organization said it would not be surprised by an estimate between 1.3x and 3x.
METR estimates that perhaps 80% of the acceleration is attributable to AI assistance in research and development. It said it was confident the change did not result from faster growth in training compute. Those are METR’s judgments, not direct measurements of a single cause. The assessment drew on benchmark scores for unreleased models, training metrics, surveys and about 15 staff interviews. Around 2.5 full-time-equivalent METR staff worked on the review using Anthropic devices; METR said it received no payment.
The report also describes Anthropic’s internal use of AI. According to Anthropic’s own measurement, Claude led 26% of its AI research and development work in August 2026, compared with less than 1% in February. Anthropic reported about 30,000 agents operating at any one time on its main internal platform. These figures show reported use and leadership of tasks; they do not establish that AI systems independently set research goals or build successor models without human direction.
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 Speed Matters
The question is whether AI can help shorten the cycle of building more capable AI systems. If that assistance materially raises the pace of progress, companies and policymakers may have less time to evaluate new capabilities and adjust safeguards. METR’s estimate is relevant because it is an outside assessment using internal company information, rather than a conclusion drawn only from public model releases.
The result is not proof of fully autonomous or runaway development. METR measures capability progress on specified tasks and estimates how much AI assistance contributed; that is different from showing that AI can independently design and develop its own successor. Still, a near-doubling estimate bears directly on Anthropic’s own early-warning rules for automated research and development. The report’s redactions also matter: a finding can be consequential while remaining difficult for independent readers to scrutinize.
Popular right now
De’Longhi Classic Matte Black Espresso Machine with Milk Frother
De'Longhi La Specialista Arte Evo Espresso Machine with Grinder
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
Anthropic’s Twofold Progress Trigger
Anthropic’s Responsible Scaling Policy, version 3.4, took effect on July 8, 2026. It identifies two triggers related to automated research and development: a model able to replace the company’s research staff at competitive cost, or a doubling in the pace of progress plausibly attributable in substantial part to automated research or engineering. The policy defines a doubling as making, in one year, the progress otherwise expected over two years.
Anthropic describes that threshold as a potential early warning, not proof that a serious risk has already materialized. In a November 2 filing with the U.S. Securities and Exchange Commission, the company said it “may imminently meet” its definition of dramatically accelerated progress. METR judged it plausible that the policy’s 2x threshold had already been crossed by July 15, but its central estimate for the measured acceleration is 1.9x. These are related assessments, not identical findings.
METR’s estimate also depends on what models and comparisons are included. When it measured frontier models across developers rather than Anthropic models alone, the estimated acceleration was smaller: from about 17 to 24 points per year, or 1.4x. METR said this could mean Anthropic lagged the frontier in 2025, making its own-company acceleration look larger, or that its earlier models were less tuned to the benchmarks, potentially making the estimate smaller.
““Perhaps ≈80%” of the current acceleration is due to AI.”
— METR
Evidence Behind the Estimate
Most of the evidence METR found most informative was redacted before publication. Anthropic supplied the data and retained the right to redact non-public material. METR also dropped a planned check of how Anthropic divides compute between safety work and capability development. The public report therefore does not allow readers to inspect all the underlying evidence or that planned comparison.
The 1.9x figure is an estimate, not a direct count of research output. METR says results shift with assumptions about task difficulty, ranging from 1.1x to 2.9x, although most of its other checks remained near 2x. The assessment concerns progress on well-specified capability tasks; it does not establish that the same pace applies to every kind of AI research or that AI systems are developing successors autonomously. The public material does not settle how much of the acceleration comes from AI assistance, nor how the redacted evidence might affect outside evaluation.
Next Tests of the Warning Threshold
The immediate next step is further scrutiny of METR’s method and the evidence that can be made public. The report’s estimate and Anthropic’s policy language give researchers a basis for comparing future measurements, but the redactions leave important details unavailable. Any later assessment will be easier to evaluate if it provides clearer evidence about benchmark selection, model performance and the role of AI tools in research work.
Anthropic’s Responsible Scaling Policy makes the 2x progress measure an early-warning trigger for additional safeguards when the acceleration is plausibly substantially attributable to automated research or engineering. Whether the company formally determines that the trigger has been met, and what measures would follow, remains unresolved in the source material. METR’s report does not itself establish that Anthropic has crossed the policy threshold.
Key Questions
What did METR find about Anthropic’s AI progress?
METR estimates that Anthropic’s capability progress on specified tasks accelerated by about 1.9x after a break point it places at October 7, 2025. It gives a 95% confidence interval of 1.5x to 2.8x and says other assumptions can produce a wider range.
Does the report show that AI is building AI on its own?
No. METR estimates that AI assistance likely accounts for much of the acceleration, but the assessment does not show fully autonomous successor development. Anthropic has said humans still choose goals and that recursive self-improvement is not inevitable.
How much of the evidence is public?
Anthropic redacted a majority of the evidence METR described as most informative. METR also did not complete a planned check of how compute is split between safety and capabilities, limiting the public’s ability to assess parts of the analysis.
Has Anthropic confirmed that it crossed its 2x threshold?
The source material does not report a final confirmation that Anthropic’s policy trigger has been met. METR said crossing it by July 15 was plausible, while Anthropic’s November 2 SEC filing said it “may imminently meet” its definition of dramatically accelerated progress.
Source: ThorstenMeyerAI.com
