🔍 Read the full analysis: Is AI Now Speeding Up Its Own Development? What METR Found Inside Anthropic, And What Stayed Redacted 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 estimates that Anthropic’s progress on measured AI capabilities accelerated about 1.9 times after October 2025, with AI use 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, limiting public scrutiny.
METR, a nonprofit that evaluates frontier AI systems, said on November 5 that Anthropic’s measured capability progress accelerated about 1.9 times after October 2025, based on 11 weeks examining internal company data. METR estimates that AI tools likely account for most of the increase, but says Anthropic redacted a majority of the evidence the evaluators found most informative, leaving important parts of the analysis difficult for outsiders to verify.
METR’s central estimate compares Anthropic’s pace of progress on a combined benchmark measure before and after an estimated turning point of October 7, 2025. The group estimates progress rose from about 15 index points per year to 29, a ratio of 1.9. METR gives a 95% confidence interval of 1.5 to 2.8 times and says it would not be surprised by estimates between 1.3 and 3 times. Its measure, AECI, combines scores across multiple capability benchmarks; it is not a direct count of research tasks completed or a measure of overall staff productivity.
METR attributes perhaps about 80% of the acceleration to AI use, describing that figure as an estimate rather than a precise accounting. The evaluators say they are confident the increase did not come from faster growth in training compute. Anthropic’s own figures, as reported in the assessment, indicate Claude led 26% of the company’s 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. Those are Anthropic measurements, not independent confirmation that agents caused the full change in benchmark progress.
The assessment used scores from unreleased models, training metrics, staff surveys and about 15 interviews. METR says roughly 2.5 full-time-equivalent staff worked on the assessment from August 12 through October 29, using Anthropic-provided laptops and free model tokens; it says it received no payment. Anthropic retained the right to redact non-public material. METR also reported that it dropped 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
Why AI-Assisted R&D Matters
The result bears on whether AI systems are beginning to speed up the work that produces more capable AI. If that effect grows, development cycles could shorten, giving companies and regulators less time to evaluate new systems and respond to risks. METR’s estimate is relevant to that debate, but it describes a measured change in benchmark progress at one company—not proof that AI can independently build its successor.
Anthropic’s own Responsible Scaling Policy sets a threshold for “automated R&D” that includes a doubling in the pace of progress plausibly attributable in substantial part to automated research or engineering. METR says it is plausible the threshold was crossed by July 15, 2026. Anthropic wrote in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. These statements indicate the issue is being treated as a potential early warning; they do not establish that fully autonomous development has arrived.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Midnight
Bose QuietComfort 45 SE Wireless Noise-Cancelling Over-Ear Headphones
Sony WH-1000XM5 Premium Noise Cancelling Wireless Headphones, Midnight Blue
As an affiliate, we earn on qualifying purchases.
How METR Measured Anthropic’s Progress
METR’s assessment differs from an external test of a finished public model: evaluators worked with internal company data, including benchmark results for models that had not been released. The central comparison uses Anthropic’s AECI capability index and an estimated October 2025 break point. METR says the measure is intended to track progress across tasks, not to imply that every part of AI research and development has sped up by the same amount.
The result depends on the comparison being made. Looking at frontier models from all developers, rather than Anthropic models alone, METR estimates progress increased from about 17 to 24 index points a year, or 1.4 times. The report offers two possible explanations: Anthropic may have been behind the frontier in 2025, which could make its own acceleration look larger, or its earlier models may have been less tuned to benchmarks, which could make it look smaller. Changing assumptions about task difficulty shifts the estimate from 1.1 to 2.9 times, while METR says most other checks remain near two times.
Anthropic’s policy defines a twofold acceleration as achieving in one year the amount of progress that would have taken two years at the earlier pace. The company has described recursive self-improvement as an AI system fully autonomously designing and developing its successor, while saying current systems are not there and that the outcome is not inevitable. The METR findings concern AI helping people with research and engineering, a narrower claim than autonomous self-improvement.
“METR estimates the acceleration at 1.9x, with a 95% confidence interval of 1.5x to 2.8x.”
— METR assessment
Evidence and Causation Still Limited
The public cannot fully inspect METR’s basis for its conclusion because Anthropic redacted a majority of the evidence the evaluators considered most informative. The source material does not specify all the withheld evidence or explain how each redaction affects the estimate. METR’s work was independent, but the company supplied equipment and tokens and had authority to redact non-public material.
The 1.9-times figure is an estimate tied to a particular index, selected models, baseline period and assumptions. It does not show that every research activity accelerated by that amount, nor does it establish that AI tools alone caused the change. METR’s estimate that AI accounts for about 80% is qualified, and its alternative all-developer comparison is lower at 1.4 times. The report also leaves the planned safety-versus-capability compute analysis undone.
It remains unclear how much the redactions prevent independent reproduction, whether Anthropic’s agent-use figures capture work that directly improved model capabilities, and whether the reported pace will persist. The source material provides no detailed public evidence on how the estimated acceleration affected safety work or the timing of particular model releases.
Policy Thresholds and Further Evidence
The next test is whether Anthropic’s own monitoring and disclosures establish that its Responsible Scaling Policy threshold has been met. Its November 2 filing said it may meet the definition imminently; METR said a crossing by July 15 was plausible. Neither statement, on the information available here, settles whether the policy trigger has been formally reached or what additional safeguards would follow.
Further public detail could help outside researchers assess the benchmark choices, uncertainty range, attribution to AI tools and effect of the redactions. METR’s findings also leave open whether other laboratories are seeing comparable changes. For now, the report is a significant measurement of one company’s internal progress, but its limited public evidence means the estimate should be read with its stated qualifications.
Key Questions
What did METR find about Anthropic’s AI progress?
METR estimated that Anthropic’s measured capability progress accelerated 1.9 times after October 2025 compared with its earlier pace. Its 95% confidence interval is 1.5 to 2.8 times, and the estimate depends on the benchmark index and assumptions used.
Does the report show that AI can build its own successor?
No. The assessment concerns AI helping people with research and engineering. Anthropic defines recursive self-improvement as a system fully autonomously designing and developing its successor, and says current systems have not reached that point.
How much of the acceleration did METR attribute to AI?
METR estimated that perhaps about 80% of the acceleration came from AI use. It presents this as an approximate judgment, not a precise causal measurement.
Why can’t readers verify the full assessment?
Anthropic redacted a majority of the evidence METR said it found most informative. The report also says Anthropic supplied laptops and free tokens and retained the right to redact non-public information; METR says it took no payment.
Did METR say Anthropic crossed its safety-policy threshold?
METR said it was plausible the threshold had been crossed by July 15, 2026. Anthropic’s November 2 filing said it may imminently meet its definition of dramatically accelerated progress. The material does not confirm a formal determination that the threshold was crossed.
Source: ThorstenMeyerAI.com
