🔍 Read the full analysis: Does AI Contribute To Its Own Development? METR On Anthropic on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s November 5 assessment estimates Anthropic’s progress on well-specified AI capability tasks is running about 1.9 times its previous pace. The nonprofit attributes much of the acceleration to AI assisting research and development, but says Anthropic redacted a majority of the evidence it considered most informative.
METR, a nonprofit that evaluates frontier AI systems, estimated on November 5 that Anthropic’s progress on measured AI capabilities is running at about 1.9 times its earlier pace, with AI assistance likely accounting for most of the acceleration. The assessment draws on 11 weeks of access to internal company data, but METR says Anthropic redacted a majority of the evidence it found most informative, limiting what outsiders can independently verify.
METR compared Anthropic’s model capability scores before and after an estimated change point of October 7, 2025. Its measure, AECI, combines results across multiple benchmarks. The reported trend rose from about 15 points per year before the change point to about 29 points per year afterward. METR describes that as roughly two years of historical progress occurring in one year. Its estimate of the acceleration was 1.9x, with a 95% confidence interval of 1.5x to 2.8x.
The estimate is about the pace of progress on the measured tasks, not a claim that every part of AI development is twice as fast or that an AI system independently creates its successor. METR said it would not be surprised by a result between 1.3x and 3x. Under different assumptions about task difficulty, its estimate ranged from 1.1x to 2.9x. METR also found a smaller increase when comparing frontier models across developers: about 1.4x, from 17 to 24 points per year.
On the cause, METR said it was confident the acceleration did not come from faster growth in training compute. It estimated that perhaps about 80% was attributable to AI, while presenting that figure as an estimate rather than a directly measured share. The report describes AI systems helping researchers and engineers, with humans still setting goals and directing work. Anthropic’s own figures, as reported in the source material, said Claude led 26% of its AI research and development work in August 2026, compared with under 1% in February, and that roughly 30,000 agents were active at a time on its main internal platform.
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 matters
If AI tools help produce improved models, they could increase the rate at which AI capabilities advance. That matters to safety researchers and policymakers because safeguards, evaluation methods and governance processes may need to keep pace with changes in model capability. METR’s finding is relevant to that debate, but it does not establish that fully autonomous recursive self-improvement has begun.
The result also bears on Anthropic’s own safety commitments. Its Responsible Scaling Policy identifies a doubling in the pace of progress, plausibly attributable to automated research or engineering, as one trigger for action around automated R&D. METR said it was plausible that the company’s early-warning threshold had been crossed by July 15. Anthropic’s November 2 SEC filing said the company “may imminently meet” its definition of dramatically accelerated progress. These statements describe a possible policy threshold, not confirmation that a specific risk has materialized.
For readers, the report is notable both for its estimate and for the conditions under which it was produced. It offers an outside assessment based on internal evidence, but the extent of redactions makes the findings harder to check independently. That tension is central: access to company data can improve evaluation, while restrictions on publication can limit public scrutiny.
Products Worth Considering
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
Bose SoundLink Plus Portable Bluetooth Speaker, Black
DJI Mic 3 Wireless Microphone Kit (2 Transmitters, 1 Receiver, Charging Case)
As an affiliate, we earn on qualifying purchases.
How METR assessed Anthropic
METR’s work differs from a typical external evaluation, which tests a released model. According to the source material, about 2.5 full-time-equivalent staff worked on the assessment from August 12 to October 29, 2026, using Anthropic devices and internal materials. These included benchmark results for unreleased models, training metrics, surveys and about 15 interviews with company staff. METR said it received no payment; Anthropic supplied laptops and free model tokens and retained the right to redact non-public information.
The report’s comparison depends on its selected measure and baseline. METR placed the change point in October 2025 and compared Anthropic’s capability-score trends on either side. The source material says the 1.9x result is higher than a 1.7x figure in Anthropic’s Haiku 5.5 system card, in part because METR included newer, unreleased models. The cross-developer estimate of 1.4x provides another comparison, but does not resolve whether Anthropic’s earlier position relative to the frontier or differences in benchmark tuning affect the company-specific result.
Anthropic’s policy defines accelerated automated R&D partly through a doubling in the pace of progress that can plausibly be attributed to research or engineering automation. It also lists a separate trigger involving a model capable of replacing the company’s research staff at competitive cost. Anthropic’s policy describes the pace threshold as a potential early warning; it is not, by itself, proof of fully autonomous development.
““perhaps ≈80%””
— METR
Redactions limit independent checks
The public cannot see much of the material METR considered most informative: the assessment says Anthropic redacted a majority of that evidence. The source material also says a planned examination of how the company divides compute between safety work and capability development was dropped. It does not provide enough detail to establish how those omissions affect the result.
The 1.9x figure is an estimate, not a direct count of development speed across every part of the company. Results vary with the benchmark set, comparison group and assumptions about task difficulty. METR’s estimate that AI accounts for roughly 80% of the acceleration is also an attribution judgment, not a precise measurement of causation. The source material does not include a full public record of the underlying data or Anthropic’s detailed response to each finding.
It is also unclear whether the estimated acceleration will continue, how broadly it applies beyond the selected tasks, or whether Anthropic has formally determined that its policy trigger has been crossed. The available information does not establish that an AI system can independently direct its own successor’s development.
Policy decisions and further evidence
The next concrete question is whether Anthropic formally concludes that it has met the automated R&D threshold in its Responsible Scaling Policy and what safeguards would follow. Its November 2 filing said the threshold may be imminent, while METR judged that crossing it by July 15 was plausible. Neither statement, in the source material, confirms a final determination.
Further scrutiny will depend on whether more evidence can be made public or assessed by independent evaluators under conditions that preserve meaningful review. Readers should also watch for additional model results and updated measurements: a single estimate, especially one with extensive redactions and a wide uncertainty range, cannot show whether the pace is stable or changing.
Key Questions
What did METR conclude about Anthropic’s AI development pace?
METR estimated that Anthropic’s progress on measured capability tasks accelerated by about 1.9x compared with its earlier rate. It reported a 95% confidence interval of 1.5x to 2.8x.
Does the report say AI is building the next AI on its own?
No. The account describes AI assisting human research and engineering. It does not establish that an AI system independently sets goals or develops its successor without human direction.
How much of the acceleration did METR attribute to AI?
METR estimated that perhaps about 80% was due to AI assistance. The figure is an estimate, not a precise causal measurement.
Why is the estimate difficult to verify?
METR said Anthropic redacted a majority of the evidence it found most informative. The report also relies on choices about benchmarks, baselines and task difficulty, which affect the estimated rate.
Has Anthropic confirmed that its policy threshold was crossed?
The source material does not report a definitive confirmation. Anthropic said in a November 2 SEC filing that it “may imminently meet” its definition, while METR said a crossing by July 15 was plausible.
Source: ThorstenMeyerAI.com
