TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A public account of a METR investigation says Anthropic’s capabilities progress ran about twice as fast from October 2025 to October 2026 as in the prior 12 months. The account’s author says AI automation is probably the main reason the pace has continued, but much of the supporting evidence is redacted and the estimate is uncertain.
A public account of a METR investigation says Anthropic’s AI capabilities progress ran at about twice its previous pace between October 2025 and October 2026, with AI automation likely responsible for most of the sustained acceleration. The finding could indicate that AI systems are helping speed up the research that builds more capable AI, though much of the underlying evidence has been redacted and the account’s author did not work on the investigation.
The account, published by the author of Planned Obsolescence, describes a METR report based on a deep-access investigation into how much Anthropic’s AI systems are accelerating its research and development. The reported comparison is between capabilities progress from October 2025 to October 2026 and progress over the preceding 12 months. The measure was an internal metric called the Anthropic Epoch Capabilities Index, or AECI.
The author says the report’s authors concluded that AI automation was probably the large majority of the reason progress remained faster, even if several factors may have contributed to the initial pickup. The source account says other supporting evidence was redacted, including what the report authors described as a majority of the evidence they found most informative. The public account does not provide the underlying data or methods in enough detail for independent readers to reproduce the assessment.
Anthropic stated in its S1 that it “may imminently meet” the threshold for dramatically accelerated progress under its Responsible Scaling Policy. The policy, effective July 2026, defines that threshold as overall capabilities progress reaching or being expected to reach twice the fastest sustained rate from before AI contributed substantially to AI research and development, with the acceleration plausibly attributable to AI automation rather than increased compute or other inputs. The account’s author interprets the report as indicating that the threshold has already been crossed; that is the author’s reading, not an independently established conclusion in the material provided.
How AI Could Speed Up AI Research
The reported finding matters because it concerns a potential feedback loop: AI tools help researchers develop better AI systems, which may then contribute more effectively to further research. If that cycle is occurring at scale, progress could outpace forecasts that assume human research labor remains the main limiting factor. The report’s result is about Anthropic’s capabilities progress, however, and does not establish that the same acceleration applies across the AI industry.
The account’s author estimates that doubling overall progress through software advances could require a much larger increase in researcher productivity. Based on assumptions about the roles of research labor and computing resources, the author puts the initial labor uplift at roughly four to nine times, and suggests it may now be around 10 times as diminishing returns make new advances harder to find. Those are model-based estimates, not figures directly measured by the report. The account also cites informal staff self-reports of about 12 times uplift, which are not presented as a formal or representative measurement.
If AI assistance continues to rise, it could affect how quickly companies reach new capability levels and how much time outside researchers have to evaluate risks. But the size and durability of any effect depend on evidence that is not public, as well as factors such as computing capacity, the difficulty of finding useful research ideas, and whether AI-generated work translates into measurable capability gains.
Products Worth Considering
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
As an affiliate, we earn on qualifying purchases.
Anthropic’s Policy Threshold
Anthropic’s Responsible Scaling Policy sets out conditions under which the company would classify progress as dramatically accelerated because of AI. As described in the source account, the benchmark compares overall capabilities progress with the fastest sustained pace before AI made a substantial contribution to AI research. The policy also requires a plausible link to AI automation, rather than acceleration from more compute or other inputs.
The public account is not the METR report itself. Its author says they did not work on the project and did not see the unredacted report; they are interpreting the public material and their reading of Anthropic’s policy. That distinction matters: the reported twofold pace and the author’s judgment that the policy threshold has been crossed should not be treated as a fully public, independently verifiable audit.
The account also explains why researcher productivity cannot be read directly from the overall capabilities rate. Researchers generate and test ideas, and testing requires computing resources; software improvements are only one input into system performance. The author uses a simplified production model to estimate the labor uplift needed, while warning that the assumptions and resulting figures are uncertain.
Evidence Behind the Acceleration
The public material leaves key questions unresolved. The report’s additional evidence and the reasons for attributing most of the sustained acceleration to AI automation are redacted, and the source account does not include enough information to independently check the AECI comparison. It is also unclear precisely how the metric is constructed, how uncertainty around its estimates was handled, or whether the result has been reviewed or replicated by outside researchers.
The account’s estimates of a roughly four-to-ninefold initial labor uplift and a possible 10-fold uplift now rely on simplified assumptions about how human labor, experimental compute, and software advances contribute to capabilities. The cited 12-fold figure comes from informal staff self-reports, not a described systematic survey. The available material also does not establish whether faster progress will persist, spread to other organizations, or translate into a comparable acceleration in every area of AI development.
More Public Detail Needed
The next useful development would be publication of more of the METR report’s methods, evidence, and uncertainty estimates, subject to any confidentiality limits. That would let researchers assess the AECI comparison and the claim that AI automation, rather than compute growth or other inputs, explains most of the sustained change.
Anthropic’s S1 language indicates that the company was assessing whether it might meet its policy threshold, but the source material does not state what formal action or public announcement will follow. Until further evidence is available, the reported acceleration is a significant signal about one company’s research pace, not a settled forecast of how quickly the entire field will advance.
Key Questions
What did the METR report reportedly find?
The public account says Anthropic’s capabilities progress was about twice as fast from October 2025 to October 2026 as during the previous 12 months, using an internal metric called AECI.
Is it confirmed that AI caused the acceleration?
The account says the report authors judged AI automation probably responsible for most of the sustained acceleration. Much of their supporting evidence is redacted, so the public material does not allow independent verification of that attribution.
Does the finding apply to all AI companies?
No. The reported investigation concerns Anthropic. The source material does not establish that other companies have experienced the same pace or cause of acceleration.
What does Anthropic’s acceleration threshold mean?
As summarized in the account, the Responsible Scaling Policy sets a benchmark of overall capabilities progress reaching twice the fastest sustained rate from before AI contributed substantially to AI research, with the change plausibly attributable to AI automation rather than increased compute or other inputs.
How certain are the estimates of researcher productivity gains?
They are uncertain. The account’s estimates of roughly four to nine times initially and possibly about 10 times now come from a simplified model, while the cited 12-times figure is described as informal staff self-reporting rather than a formal measurement.
Source: planned-obsolescence.org
