🔍 Read the full analysis: Is AI Speeding Up Its Own Development? METR’s Anthropic Findings on ThorstenMeyerAI.com
Get movie nights 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 estimates that Anthropic’s progress on well-specified AI tasks accelerated about 1.9 times after October 2025, with AI assistance likely accounting for much of the increase. The nonprofit says Anthropic redacted a majority of the evidence it considered most informative, leaving important parts of its assessment difficult for outsiders to verify.
METR, an independent AI evaluation nonprofit, estimates that Anthropic’s progress on well-specified AI tasks accelerated to about 1.9 times its earlier pace, according to an assessment published November 5 after 11 weeks working with internal company data. METR attributes much of the increase to AI systems helping with AI research and development, but says Anthropic redacted a majority of the evidence it found most informative, limiting what outsiders can independently inspect.
METR compared Anthropic’s capability progress before and after a change point it estimates occurred on October 7, 2025. Its measure, AECI, combines performance across multiple benchmarks. METR estimates the score was rising by about 15 points per year before the change and about 29 points per year afterward. The resulting estimate is 1.9 times the earlier rate, with a reported 95% confidence interval of 1.5 to 2.8 times.
The organization says the estimate is sensitive to assumptions. When it considered the frontier models of multiple developers rather than Anthropic alone, the estimated acceleration was 1.4 times, from about 17 to 24 points per year. METR says different assumptions about task difficulty move its Anthropic estimate between 1.1 and 2.9 times, while most other checks it describes remain near twice the earlier pace.
METR estimates that perhaps about 80% of the acceleration may be attributable to AI use, while saying it is confident faster growth in training compute did not cause the increase. That is an estimate, not a directly observed allocation of causation. Anthropic’s own measurement, cited in the source report, says Claude led 26% of its AI research and development work in August 2026, compared with less than 1% in February. The company also reported about 30,000 agents operating at a time on its main internal platform.
METR said about 2.5 full-time-equivalent staff worked on the assessment using Anthropic devices, internal benchmark and training information, surveys and roughly 15 interviews with staff. METR says it took no payment. Anthropic supplied laptops and free tokens and retained the right to redact non-public information. The assessment also did not include a planned examination of how Anthropic divides compute between safety work and capability development.
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
What Faster AI Research Could Mean
If AI tools help researchers produce better models more quickly, the effect could compound: each generation may assist work on later generations. METR’s result is relevant to how quickly AI capability development could advance, and to whether safety practices and oversight can keep pace. The estimate does not establish that an AI system is autonomously designing and building its own successor.
Anthropic’s Responsible Scaling Policy sets a threshold for “automated R&D” that includes a doubling of progress plausibly attributable in substantial part to automated research or engineering. METR says it is plausible that the company crossed its early-warning threshold by July 15. The source material also reports that Anthropic said in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. These are assessments against a company policy, not proof that fully autonomous self-improvement is underway.
The limited public evidence matters because readers, researchers and policymakers cannot fully evaluate the result from a headline figure alone. The estimate points to a possible acceleration, while the redactions restrict independent scrutiny of how it was calculated and what information could change the interpretation.
Popular right now
De'Longhi La Specialista Touch Semi-Automatic Espresso Machine with Grinder
Dyson Ball Animal 3 Upright Vacuum – Corded, 290 AW, De‑tangling Motorbar Cleaner Head, 3 Suction Modes
iRobot Roomba 105 Combo Robot Vacuum and Mop
As an affiliate, we earn on qualifying purchases.
How METR Measured Anthropic’s Progress
METR’s work differs from an evaluation that tests only a publicly released model. The nonprofit worked with unreleased model benchmark scores, training metrics, staff surveys and interviews, according to the supplied account. Its researchers worked on Anthropic premises and devices from August 12 through October 29, 2026. The result is an external assessment built from internal material, but not one whose full evidentiary basis is public.
Anthropic’s Responsible Scaling Policy version 3.4 took effect July 8, 2026, according to the source. It identifies two relevant kinds of automated R&D trigger: a model able to replace the company’s research staff at competitive cost, or a doubling in the rate of progress plausibly substantially attributable to automated research or engineering. The policy describes the doubling threshold as achieving in one year the progress previously expected over two.
METR’s 1.9-times estimate is higher than the 1.7-times figure in Anthropic’s Haiku 5.5 system card. METR says including newer, unreleased models is the main reason for the difference. It also cautions that a company-only trend may be affected by whether Anthropic’s earlier models lagged the wider frontier or were less tuned to benchmarks.
““We are not there yet, and recursive self-improvement is not inevitable.””
— Anthropic
What the Redactions Leave Unknown
METR says Anthropic redacted a majority of the evidence it found most informative. The supplied account does not identify every withheld item or explain how each redaction affects the report’s conclusions. That limits outside readers’ ability to reproduce the analysis or assess the weight of individual data sources.
The 1.9-times figure is an estimate based on a particular benchmark index, time series and set of modeling choices. The wider-frontier estimate is lower, at 1.4 times, and the estimate changes under alternative assumptions about task difficulty. The report therefore does not establish a single, precisely known rate of acceleration across the entire AI sector.
It is also unclear from the available material how much of the change reflects AI assistance versus other factors in research workflows, or how the excluded compute-allocation analysis might affect understanding of the balance between safety and capability work. Claude’s reported share of work that it “leads” does not, by itself, show how much work it completed, how that measure is defined, or how much it changed research output.
Further Evidence Needed to Verify the Trend
The next step for evaluating METR’s conclusion is greater public clarity about the data and methods that could be disclosed without exposing sensitive company information. The source material does not announce a date for additional releases, nor does it specify that Anthropic plans to publish the redacted evidence or restore the omitted compute-allocation check.
Readers should also watch for updates to Anthropic’s Responsible Scaling Policy assessments and future evaluations using comparable benchmarks. Repeated measurements could help show whether the estimated increase persists, whether the trend extends beyond Anthropic, and how much of it is linked to AI-assisted research. Until then, the reported acceleration remains a consequential but partly unverifiable estimate, not confirmation of fully autonomous recursive self-improvement.
Key Questions
What did METR find about Anthropic’s AI development pace?
METR estimates that Anthropic’s progress on well-specified tasks accelerated to about 1.9 times its earlier pace after a change point it places in October 2025. Its 95% confidence interval is 1.5 to 2.8 times.
Does the report show that AI is building the next AI by itself?
No. METR’s assessment concerns AI assistance with research and development. Anthropic describes fully autonomous design and development of a successor as a separate condition and says it is not there yet.
How much of the acceleration does METR attribute to AI?
METR estimates that perhaps about 80% of the acceleration is due to AI, but this is an estimate. The report does not establish that share as a directly measured fact.
Why is the estimate difficult to check independently?
METR says Anthropic redacted a majority of the evidence it found most informative. The assessment used internal data, including unreleased-model benchmarks, and the public record does not provide all of the material needed to reproduce the analysis.
Did METR find the same acceleration across the AI industry?
No industry-wide doubling is established by the figures provided. When METR examined frontier models across developers, it estimated a smaller 1.4-times increase in the pace of progress.
Source: ThorstenMeyerAI.com
