🔍 Read the full analysis: How METR Investigated AI’s Role In Its Own Development 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 independent assessment estimates Anthropic improved its models on well-specified tasks about 1.9 times faster after October 2025 than before, and says AI use likely accounts for most of the acceleration. The report draws on internal company data but says Anthropic redacted much of the evidence METR considered most informative, leaving important parts of the finding difficult for outsiders to verify.
METR, a nonprofit that evaluates frontier AI systems, estimates that Anthropic’s progress on well-specified AI capability tasks accelerated about 1.9 times after October 2025, compared with the company’s earlier pace. The assessment, published November 5 after researchers examined internal company data, attributes most of the estimated speed-up to AI helping with AI research and development; METR also says much of its most informative supporting evidence was redacted.
METR’s central estimate compares Anthropic’s rate of capability progress before and after a change point it places at October 7, 2025. Using a measure based on benchmark scores, METR estimates progress rose from about 15 index points per year to 29. In METR’s framing, that is roughly two years of earlier progress achieved in one year. The estimate is not a direct measure of how quickly every part of AI development now runs.
The report says the acceleration is unlikely to be explained by faster growth in training compute. METR’s estimate is that perhaps 80% of the speed-up came from AI use, but that figure is an attribution judgment, not a directly measured share. The report gives a 95% confidence interval of 1.5x to 2.8x for the acceleration estimate and says it would not be surprised by a result between 1.3x and 3x.
METR’s work drew on unreleased-model benchmark results, training metrics, surveys and about 15 interviews with Anthropic staff. Around 2.5 full-time-equivalent researchers worked on the assessment over 11 weeks, using Anthropic devices. METR says it received no payment; Anthropic supplied laptops and free model tokens and retained the right to redact non-public information.
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
How Faster AI Research Could Affect Safety
The finding matters because AI-assisted research could shorten the time between generations of AI systems. If that assistance continues to grow, labs may have less time to evaluate new capabilities and put safeguards in place before another round of development. METR’s estimate points to a possible acceleration, but it does not establish that AI systems are independently designing their successors or that a rapid, self-sustaining cycle has begun.
Anthropic’s own Responsible Scaling Policy includes a threshold for automated research and development: progress comparable to two years at the prior rate in a single year, when substantially attributable to research or engineering automation. METR says it is plausible that this early-warning threshold had been reached by July 15, 2026. That is METR’s assessment, not a confirmed determination that Anthropic’s policy trigger was met.
The public record has a significant constraint: METR says Anthropic redacted a majority of the evidence it found most informative. Readers can assess the reported method and headline estimate, but cannot independently inspect much of the material behind it.
Products Worth Considering
Apple MacBook Pro MPHF3LL/A 14.2-Inch Laptop | M2 Pro 12-Core Chip 16GB Unified Memory 1TB SSD | 19-Core GPU Liquid Retina XDR Display Space
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, Touch ID, Wi-Fi 7, Midnight
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p FaceTime HD Camera – Blush
As an affiliate, we earn on qualifying purchases.
How METR Assessed Anthropic
This was not a conventional evaluation in which an outside group tests a publicly available model. METR instead worked with internal company data, including results for models Anthropic had not released. Its researchers also gathered staff surveys and conducted interviews, giving the assessment access to information that would not usually be visible to external evaluators.
METR used AECI, Anthropic’s version of the Epoch Capabilities Index, to combine scores across multiple benchmarks into a capability measure. Its 1.9x estimate is higher than the 1.7x figure in Anthropic’s Haiku 5.5 system card; METR says including newer, unreleased models is a main reason for the difference. When it compared frontier models across developers rather than Anthropic alone, the estimated change was smaller: from about 17 to 24 points per year, or 1.4x.
These estimates concern capability progress on well-specified tasks, not a general measure of research productivity. METR says changing assumptions about task difficulty can move its estimate from 1.1x to 2.9x, although most of its other checks remained near 2x. Anthropic’s policy describes recursive self-improvement as a system autonomously designing and developing its successor; the source report says the current situation instead involves AI assisting human researchers, who continue to choose goals.
“perhaps ≈80%”
— METR
What the Redactions Leave Unknown
The assessment is constrained by the material the public cannot see. METR says Anthropic redacted a majority of the evidence it found most informative and removed a planned examination of how the company divides compute between safety work and capability development. The available account does not give readers enough detail to independently reproduce the full analysis.
The 1.9x estimate also depends on the index, the models included and assumptions about task difficulty. METR offers alternative explanations for why its Anthropic-only estimate is higher than its all-developer comparison: Anthropic may have been behind the frontier in 2025, which could inflate the apparent acceleration, or its earlier models may have been less tuned to benchmarks, which could make the estimate too low. The source does not resolve between those possibilities.
Nor does the report establish that Anthropic’s policy threshold was formally triggered. METR says a crossing by July 15 is plausible. Separately, an Anthropic SEC filing dated November 2 said the company “may imminently meet” its definition of dramatically accelerated progress. The filing wording is prospective and qualified; it is not confirmation that the threshold had been met.
Watch for Policy and Evidence Updates
The next points to watch are whether Anthropic publicly clarifies its assessment under its Responsible Scaling Policy and whether additional evidence from METR’s work can be released. The source material does not specify a date for either development. Anthropic’s November 2 filing indicates it was considering whether it might imminently meet its own definition of dramatically accelerated progress, but does not confirm a decision or policy response.
Further evaluations could test whether the estimated pace persists, compare results across more companies and models, or make the connection between AI use and capability gains more measurable. Until more supporting data is available, the 1.9x figure remains an external estimate with material evidence withheld, rather than a fully public finding that readers can independently verify.
Key Questions
What did METR estimate?
METR estimated that Anthropic’s progress on well-specified AI capability tasks accelerated about 1.9 times after a change point it places at October 7, 2025, compared with its earlier rate.
Does the report show that AI is building AI autonomously?
No. The report describes AI assisting human research and development. It does not establish that an AI system is autonomously designing and developing its own successor.
How much of the acceleration did METR attribute to AI?
METR said perhaps 80% of the current acceleration is due to AI. That is an approximate attribution in the report, not a directly measured proportion.
Why is the estimate difficult to verify?
METR says Anthropic redacted a majority of the evidence it found most informative. The report also notes that estimates vary with assumptions about task difficulty and which models are compared.
Did Anthropic confirm that its policy threshold was crossed?
No such confirmation appears in the source material. METR said a crossing by July 15 was plausible, while Anthropic’s November 2 filing said it “may imminently meet” its definition.
Source: ThorstenMeyerAI.com
