🔍 Read the full analysis: AI Speeding Up AI: What METR Found—and What Stayed Redacted 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 measured capability progress is running at about 1.9 times its rate before October 2025. METR attributes perhaps 80% of the acceleration to AI use, but Anthropic redacted much of the evidence and the estimate has substantial uncertainty.
METR, a nonprofit that evaluates frontier AI systems, estimated on November 5 that Anthropic’s measured AI capability progress has accelerated about 1.9 times since October 2025, with AI assistance likely accounting for most of the increase. The report draws on 11 weeks of access to internal company data, but METR said Anthropic redacted a majority of the evidence it found most informative, limiting what outsiders can independently verify.
METR’s main estimate compares Anthropic’s progress on a composite of benchmark results before and after a change point it places at October 7, 2025. The assessment puts the earlier rate at about 15 points per year on the index and the later rate at about 29 points, producing an estimated 1.9x acceleration. METR’s 95% confidence interval is 1.5x to 2.8x; the organization said it would not be surprised by a range from 1.3x to 3x.
That result is an estimate of progress on measured tasks, not a finding that AI systems are independently designing and building their successors. METR says perhaps 80% of the acceleration may come from AI use and says it is confident the change did not result from faster growth in training compute. Those are the evaluator’s judgments, not a directly measured breakdown of causes.
The assessment also cites Anthropic’s own internal measurement: the company reported that Claude “leads” 26% of its AI research and development work in August 2026, up from under 1% in February. The source report says about 30,000 agents were working at a time on Anthropic’s main internal platform. The meaning and scope of “leads” are not fully detailed in the public material, so the figures should not be read as showing that AI performs 26% of all research independently.
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
If the estimate is broadly right, AI tools are helping researchers produce more capability progress in a given period. That could shorten the time available to evaluate new systems and respond to changes in their abilities. The finding is relevant to safety planning because Anthropic’s own policy uses a possible doubling in progress as an early-warning threshold for automated research and development.
The number does not establish that such a threshold has been met, or that AI systems are autonomously improving themselves. METR estimates a change in benchmark progress and offers a view about its likely causes. The distinction matters: faster work with AI assistance is not the same as a system independently setting research goals, developing a successor and deploying it.
The report’s redactions also matter for public oversight. Readers can see the headline estimate and some of its methodology, but cannot examine much of the underlying evidence that METR considered most informative. That makes the assessment valuable as an independent account of private data, while leaving important parts of its reasoning difficult to check from outside.
Products Worth Considering
Samsung Galaxy Tab A11 8.7-Inch Wi-Fi Tablet, 64GB, 4GB RAM, Silver
Bose QuietComfort Wireless Noise Cancelling Headphones, Black
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
How METR Built Its Estimate
METR’s work differs from an outside evaluation based only on publicly released models. From August 12 to October 29, 2026, about 2.5 full-time-equivalent staff worked with internal Anthropic data on company devices. METR says the material included benchmark scores for unreleased models, training metrics, surveys and roughly 15 interviews with key staff. The organization said it received no payment; Anthropic provided laptops and free tokens and retained the right to redact non-public information.
The central index, AECI, combines benchmark scores into a capability measure. METR’s estimate for Anthropic alone is higher than the 1.7x figure in Anthropic’s Haiku 5.5 system card, in part because METR included newer models that had not been released. Looking across frontier models from all developers instead of Anthropic alone, METR estimated a smaller rise: from about 17 to 24 points per year, or 1.4x.
METR cautions that the Anthropic-specific result can be read in more than one way. Anthropic may have lagged the frontier in 2025, which could make its later acceleration look larger; alternatively, its earlier models may have been less tuned to the benchmarks, which could make the measured acceleration smaller. With different assumptions about task difficulty, the estimate ranges from 1.1x to 2.9x, although METR said most other checks stayed near 2x.
“A majority of the evidence we found most informative.”
— METR
What the Redactions Leave Unclear
The public assessment does not provide most of the evidence METR considered most informative, and the source material does not specify every redacted item. That makes it difficult for outside researchers to independently reproduce the analysis or judge how strongly the evidence supports the proposed causes of acceleration.
The 1.9x figure is also sensitive to methodological choices and the selection of models and benchmarks. METR’s estimates differ when it looks at all frontier developers, and its task-difficulty assumptions produce a range from 1.1x to 2.9x. The report therefore supports a finding of faster measured progress under its chosen approach, not a precise, universally applicable rate for AI research as a whole.
It is also unclear from the supplied material exactly how Anthropic calculates the claim that Claude “leads” 26% of its research and development work, what tasks count toward that measure, and how much human review remains involved. METR’s estimate that AI explains perhaps 80% of the acceleration is similarly qualified. The report does not establish that fully autonomous recursive self-improvement is happening.
Next Tests of the Acceleration
The next useful evidence would include updated measurements that show whether the estimated pace persists, along with clearer definitions of Anthropic’s internal “leads” metric and more information about the redacted findings. Publicly released model evaluations could help test whether the trend extends beyond the internal data METR reviewed.
Anthropic’s Responsible Scaling Policy sets out a doubling of progress plausibly attributable to automated research or engineering as one trigger for safeguards. The source report says METR considered it plausible that Anthropic’s early-warning threshold had been crossed by July 15, while Anthropic stated in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. Whether the policy trigger has been formally met, and what action would follow, remains unresolved in the available material.
Key Questions
What did METR estimate?
METR estimated that Anthropic’s progress on its measured capability index rose from about 15 points per year before October 2025 to about 29 afterward, or 1.9 times the earlier rate. Its 95% confidence interval is 1.5x to 2.8x.
Does the report show that AI is building its own successor?
No. The report concerns AI assistance in research and development. It does not establish that an AI system is autonomously setting goals and designing and developing a successor without human direction.
Why is the estimate not fully verifiable?
METR said Anthropic redacted a majority of the evidence it found most informative. The estimate also changes under different model selections and assumptions, leaving outside researchers with limited ability to reproduce the full analysis.
What does the 26% figure mean?
Anthropic’s own measurement, cited in the report, says Claude “leads” 26% of its AI research and development work in August 2026, compared with under 1% in February. The public material does not fully define “leads,” so the figure does not show that Claude independently performs that share of all research.
Source: ThorstenMeyerAI.com
