🔍 Read the full analysis: The Anthropic Findings That Put AI Development Under The Microscope 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 progress on measured AI capabilities accelerated about 1.9 times after October 2025. The nonprofit says AI use likely explains much of the change, but Anthropic redacted most of the evidence METR considered most informative, leaving key details difficult for outsiders to verify.
METR, a nonprofit that evaluates frontier AI systems, estimated on November 5 that Anthropic’s progress on measured AI capabilities accelerated about 1.9 times after October 2025, based on 11 weeks of access to company data. METR believes AI assistance likely accounts for much of the increase, but says Anthropic redacted a majority of the evidence it found most informative, limiting what the public can independently assess.
METR’s central estimate compares Anthropic’s pace of capability improvement after an estimated October 7, 2025, change point with its earlier pace. Using a composite benchmark measure, the report estimates progress rose from about 15 points per year to 29 points per year. METR describes the result as roughly two years of historical progress occurring in one year. Its estimate of the acceleration is 1.9x, with a 95% confidence interval of 1.5x to 2.8x; the organization says it would not be surprised by a result between 1.3x and 3x.
The assessment used unreleased-model benchmark scores, training metrics, staff surveys and about 15 interviews, according to the report. About 2.5 full-time-equivalent METR staff worked on it from August 12 to October 29, using Anthropic devices. METR says it received no payment; Anthropic supplied laptops and free tokens and retained the right to redact non-public information. The report also says METR abandoned a planned analysis of how Anthropic divides compute between safety work and capability development.
The estimate changes with the comparison set. Looking across frontier models from multiple developers, rather than Anthropic alone, METR calculates a smaller acceleration, from about 17 to 24 points per year, or 1.4x. METR says the difference could mean Anthropic lagged the frontier in 2025, making its company-specific increase look larger, or that its earlier models were less tuned to benchmarks, which could make the increase look smaller. Alternative assumptions about task difficulty put the estimate between 1.1x and 2.9x.
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
AI Assistance and Anthropic’s Safety Threshold
The report addresses a consequential question for AI safety: whether AI systems are helping researchers build more capable successors quickly enough to accelerate the development process itself. METR estimates that perhaps about 80% of the acceleration it measured is attributable to AI use. It also says it is confident the change did not result from faster growth in training compute. That attribution is METR’s interpretation of the evidence, not a directly observable measure of how much AI contributed to each research task.
The estimate matters because Anthropic’s Responsible Scaling Policy sets a doubling of progress plausibly attributable to automated research or engineering as one trigger for safeguards. METR says it is plausible the threshold was crossed by July 15, 2026. Anthropic stated in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. These statements point to a policy-relevant warning, not confirmation that fully autonomous AI research or recursive self-improvement has arrived.
If AI tools help researchers complete more of the work involved in developing models, the pace of capability improvement could affect how quickly labs and regulators need to evaluate safeguards. But benchmark progress is not the same as a doubling of every part of research and development: experiments, training runs and human decisions can still impose limits. METR’s findings inform that debate; they do not establish that AI can independently design and build its successor.
Products Worth Considering
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
LG UltraGear 32GS60QC-B 32-inch Curved QHD Gaming Monitor (180Hz, 1ms, 1000R, AMD FreeSync, HDR10)
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Black
As an affiliate, we earn on qualifying purchases.
How METR Measured Anthropic’s Progress
METR’s measure, called AECI, combines scores from multiple capability benchmarks into a single index. The report compares the rate of change in that index before and after its estimated October 2025 break point. METR says Anthropic’s company-specific estimate is higher than the 1.7x figure in Anthropic’s Haiku 5.5 system card, mainly because METR included newer, unreleased models.
Anthropic’s Responsible Scaling Policy version 3.4 took effect July 8, 2026. It identifies two possible triggers related to automated research and development: a model capable of replacing Anthropic’s entire research staff at competitive cost, or progress doubling in a way plausibly attributable in substantial part to automated research or engineering. Anthropic describes the threshold as an early warning, not proof that a danger is already present, and says recursive self-improvement is not inevitable.
Anthropic also reported that Claude led 26% of its AI research and development work in August 2026, compared with under 1% in February, and that around 30,000 agents were working at a time on its main internal platform. Those are company-reported figures, not independently verified findings in the same sense as METR’s benchmark analysis. The source report says METR’s assessment was neither sponsored nor reviewed by Anthropic or METR’s own subject company; its author disclosed using Claude to help draft the analysis.
““A 2x acceleration means that 2 years of historical progress is made in 1 year.””
— METR
Redactions Limit Independent Scrutiny
The public cannot fully evaluate METR’s reasoning because, according to the assessment, Anthropic redacted a majority of the evidence METR found most informative. The report’s access conditions also gave Anthropic control over disclosures of non-public material. METR says it had internal data and staff interviews, but readers cannot inspect much of the underlying evidence described in the assessment.
The precise amount of acceleration remains sensitive to methodological choices. METR’s estimates range from 1.1x to 2.9x under different assumptions about task difficulty, while the all-developer comparison yields 1.4x rather than 1.9x. The report also leaves open how much of the apparent change reflects AI assistance, benchmark tuning, Anthropic’s position relative to other labs, or other factors. The assessment does not establish that AI systems are independently conducting the full research-and-development process, and its planned analysis of compute allocation between safety and capabilities was not completed.
Policy Decisions and Further Evidence
The immediate next step is whether Anthropic determines that it has met the automated-R&D trigger in its Responsible Scaling Policy and what safeguards it would apply if it does. The company’s November 2 filing says the threshold may be imminent, while METR’s assessment says an earlier crossing by July 15 is plausible; neither statement, in the material provided, confirms a final policy determination.
Further scrutiny will depend on whether additional evidence becomes public, including material behind the redactions and clearer information about how AI is used across research and engineering. Future assessments can also test whether the measured acceleration persists and whether comparable changes appear across other AI developers. Until then, the 1.9x estimate is a significant but qualified finding, with its interpretation constrained by the evidence readers cannot see.
Key Questions
What did METR find about Anthropic’s AI progress?
METR estimated that Anthropic’s measured capability progress accelerated 1.9x after an estimated October 2025 change point, with a 95% confidence interval of 1.5x to 2.8x. Its measure tracks benchmark capability progress, not every part of AI development.
Does the assessment show that AI is building AI autonomously?
No. METR says AI likely contributed substantially to faster progress, but the material describes AI helping people develop models. It does not establish that an AI system can autonomously design and develop its own successor.
How much of the acceleration does METR attribute to AI?
METR estimates that perhaps about 80% of the acceleration is due to AI. That is the nonprofit’s interpretation of its evidence, not a directly measured accounting of AI’s contribution to every task.
Why is the 1.9x estimate uncertain?
METR reports that Anthropic redacted most of the evidence it found most informative. The estimate also varies with assumptions about task difficulty and the models included; METR’s comparison across developers produced a 1.4x figure.
Has Anthropic said it crossed its safety-policy threshold?
Anthropic said in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. METR said a crossing by July 15 was plausible. The source material does not confirm a final determination that the threshold has been met.
Source: ThorstenMeyerAI.com
