🔍 Read the full analysis: Inside METR’s Study Of AI-Assisted Development At Anthropic on ThorstenMeyerAI.com
Get the little things that make your day 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 measured AI capabilities accelerated about 1.9x after October 2025, with AI assistance likely accounting for most of the increase. The assessment draws on 11 weeks of access to internal data, but Anthropic redacted much of the evidence METR considered most informative.
Nonprofit evaluator METR estimates that Anthropic’s AI capability progress accelerated about 1.9x after October 2025, with AI assistance likely responsible for most of the increase. The assessment, published November 5 after 11 weeks of access to company data, offers a detailed outside look at AI helping to develop AI—but Anthropic redacted a majority of the evidence METR found most informative, limiting public scrutiny of the estimate.
METR measured changes in Anthropic’s performance across a collection of benchmarks, using a composite measure called the Anthropic Epoch Capabilities Index, or AECI. It estimates progress rose from about 15 index points per year before a break point around October 7, 2025, to about 29 points per year afterward. METR describes that change as 1.9x acceleration, with a 95% confidence interval of 1.5x to 2.8x. In plain terms, its estimate is roughly two years of the earlier measured progress in one year.
The estimate concerns progress on well-specified tasks, not a claim that all AI research or the full development cycle has become twice as fast. METR says its overall judgment is around twice the earlier rate, while noting it would not be surprised by a result between 1.3x and 3x. When it examined the frontier models of all developers rather than Anthropic alone, the estimated acceleration was smaller: 1.4x, from about 17 to 24 index points per year.
METR’s report attributes perhaps 80% of the acceleration to AI use, while expressing confidence that faster growth in training compute did not explain it. That proportion is an estimate, not a direct measurement of every cause. Anthropic’s own data, cited in the source report, indicate that Claude led 26% of its AI research and development work in August 2026, compared with less than 1% in February. The company’s main internal platform reportedly had about 30,000 agents working at a time; the available material does not define “leads” in enough detail to treat that figure as a measure of completed research.
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 AI Assistance Could Change R&D
The findings matter because AI systems helping researchers build later systems could speed up the work that produces more capable models. If that effect grows, safety testing, governance and oversight may have less time to keep pace with model development. METR’s estimate is relevant to that debate, but it is not evidence that AI systems are autonomously designing their own successors.
Anthropic’s Responsible Scaling Policy identifies a doubling of progress plausibly attributable to automated research or engineering as one trigger for its “automated R&D” safeguards. METR says it is plausible the company’s early-warning threshold had been crossed by July 15, 2026. That is METR’s interpretation of the policy and evidence, not a confirmed finding that Anthropic has formally declared the threshold met. Anthropic’s November 2 SEC filing said it “may imminently meet” its definition of dramatically accelerated progress.
The report also highlights a transparency problem. METR had access to unreleased-model benchmark scores, training metrics, surveys and staff interviews, but much of what it judged most useful was withheld from the public version. Readers can see the evaluator’s conclusions, but cannot independently inspect much of the underlying evidence.
Products Worth Considering
Garmin Forerunner 255 GPS Running Smartwatch – 46 mm, Slate Gray
Garmin Venu 3S GPS Smartwatch with AMOLED Display, Soft Gold and French Gray Silicone Band
Philips Sonicare 5950 Rechargeable Electric Toothbrush with 2 Brush Heads, Pink
As an affiliate, we earn on qualifying purchases.
How METR Built Its Estimate
METR is a nonprofit that evaluates frontier AI systems. In this assessment, it took an embedded approach rather than testing only a publicly available model. About 2.5 full-time-equivalent staff worked with internal Anthropic data from August 12 through October 29, 2026, using company devices. METR says it received no payment; Anthropic provided laptops and free tokens and retained the right to redact non-public information.
The 1.9x figure is higher than the 1.7x estimate 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 the Anthropic-only estimate has interpretive limits: the company may have been behind the overall frontier in 2025, which could make its subsequent acceleration appear larger, or earlier models may have been less tuned to benchmarks, which could make the estimate smaller. Under different assumptions about task difficulty, the result ranged from 1.1x to 2.9x, while most other checks remained near 2x.
Anthropic’s policy describes a separate trigger: a model able to replace its research staff at competitive cost, or a doubling in the pace of progress that is plausibly attributable to research or engineering automation. The policy calls the threshold an early warning, not proof that a particular danger has arrived. Anthropic has also said recursive self-improvement—an AI system fully autonomously developing its successor—has not been reached and is not inevitable.
““roughly two years of historical progress over the last year.””
— METR
What the Redactions Leave Unknown
The public report does not expose much of the material METR considered most informative: Anthropic redacted a majority of that evidence. The source material also says METR dropped a planned examination of how Anthropic divides compute between safety work and capability development. Without those details, outside readers cannot fully evaluate the inputs behind the estimate or assess that allocation.
The 1.9x result is a model-based estimate whose size varies with assumptions and the comparison set. METR’s 1.4x estimate across frontier developers is a different comparison, not a direct correction of the Anthropic-specific result. The available information does not settle how much of the measured change reflects AI assistance, other changes in research practices, model tuning, or measurement effects.
It is also unclear from the available material whether Anthropic has formally concluded that its policy threshold has been crossed, what safeguards it would apply if it did, or how “Claude leads” in a share of R&D work was defined and measured. The source account notes that METR used staff interviews and internal records, but the public summary does not provide enough detail to independently resolve these questions.
Further Evidence and Policy Decisions
The next important developments are likely to be fuller disclosure of the evidence, additional measurements over time and clarification from Anthropic about its policy assessment. METR’s estimate is a snapshot based on a review ending October 29, 2026; later data could show whether the measured pace persists, rises or falls.
Readers should also watch for a clear statement from Anthropic on whether it considers its Responsible Scaling Policy’s accelerated-progress trigger to have been met and what steps would follow. Until more underlying evidence is available, the report supports a serious but qualified conclusion: METR measured a substantial acceleration in Anthropic’s benchmark progress and believes AI assistance probably explains most of it, while the public cannot fully audit the assessment.
Key Questions
What did METR estimate?
METR estimated that Anthropic’s measured capability progress accelerated about 1.9x after a break point around October 7, 2025. Its 95% confidence interval was 1.5x to 2.8x.
Does the report say AI is building AI on its own?
No. The assessment concerns AI helping people with research and development. Anthropic says its systems have not reached the point of fully autonomously designing and developing their own successors.
How much of the acceleration did METR attribute to AI?
METR estimated that perhaps 80% of the acceleration came from AI use. The report presents this as an approximate judgment, not a precisely measured causal share.
Why can’t the public fully check the findings?
Anthropic redacted a majority of the evidence METR found most informative. METR also did not complete a planned check of how the company divides compute between safety and capabilities work.
Has Anthropic confirmed that its policy threshold was crossed?
The provided material does not confirm that Anthropic formally declared the threshold met. Its November 2 filing said it “may imminently meet” its definition, while METR said crossing the early-warning threshold by July 15 was plausible.
Source: ThorstenMeyerAI.com
