Tuesday, November 17, 2026

Top 5 This Week

Related Posts

AI nearly doubled the pace of Anthropic’s progress, says independent evaluator METR


Can AI speed up its own manufacturing? Days after Anthropic warned in its amended stock prospectus that it could soon cross its threshold for highly AI-accelerated research, an independent evaluator is providing some numbers. THE November 5, 2026the organism METR published its assessment of Anthropic’s research. His conclusion: the capabilities of the models are progressing approximately twice as fast than before October 2025, and AI would be the main explanation.

An eleven-week evaluation

METR is a nonprofit organization that studies the most advanced AI agents, with the goal of determining whether developers are at risk of losing control of them. Of August 12 to October 29, 2026its teams carried out a so-called high access evaluation at Anthropic, which mobilized the equivalent of 2.5 full-time people for more than eleven weeks.

Three questions guided the work: Has the growth of Anthropic’s internal capabilities accelerated? How much of this acceleration is due to AI? Are there bottlenecks likely to slow it down? METR specifies that it has not focused on the thresholds set in Anthropic’s security policy, and its report does not say whether the threshold mentioned in the prospectus has been crossed.

An acceleration of 1.9 times

To measure progress, METR relied on an index that aggregates the results of Anthropic’s models across numerous tests, including never-published internal models. By comparing the rate of progression before and after a break date, the body obtains an acceleration of 1.9 times since October 2025, with a confidence interval of 1.5 to 2.8 times. According to METR, Anthropic has made as much progress on these tests in one year as between October 2023 and October 2025.

Indicator METR Estimate
Acceleration of capacities since October 2025 1.9 times, between 1.5 and 2.8 times
Range according to analysis variants 1.1 to 2.9 times
Share of recent acceleration due to AI About 80%, very uncertain
Benefit felt by employees with the most recent internal model Around 12 times, data considered very fragile

METR remains cautious. Robustness analyzes give results between 1.1 and 2.9 times, and the organization would not be surprised if the true figure is closer to 1.3 or 3. Its overall judgment: Anthropic improves its results on well-defined tasks at approximately twice the rate observed between March 2024 and October 2025.

The index used is Anthropic’s version of a public indicator designed by the Epoch AI research institute, which combines scores from numerous tests. METR calculated the progression slope before and after a break date, then the ratio between the two slopes. The organization supplemented this measurement with interviews with employees, internal surveys on the productivity gain felt, and other indicators such as staff numbers, the volume of code produced or the computing power devoted to training.

Anthropic had already published, in the technical sheet of its Claude Haiku 5.5 model, an acceleration estimate of 1.7 times. METR explains the discrepancy mainly by the inclusion, in its own calculation, of never-published internal models.

AI, main explanation

The organization says it is confident on one point: acceleration doesn’t come an increase in computing power devoted to training. He estimates that around 80% of the recent acceleration is attributable to AI, while highlighting high uncertainty. Progress at the end of 2025, however, would have been influenced by other factors.

How AI helps matters as much as the number. According to METR, it accelerates research especially by assisting researchersand not yet taking care of the most important parts of model training independently. The researchers only mention “signs of life” of autonomous contributions. The employees interviewed informally estimate that the most recent internal model increases their productivity by around twelve times, a figure that METR describes as very fragile evidence.

No brakes in sight

Regarding bottlenecks, METR does not see, in the next twelve months, any significant limit linked to the progress of algorithms, to the computing power for training or experiments, nor to data. The organization, however, specifies that it is not confident in this conclusion, its interpretation having changed several times during the evaluation.

It also identifies risks in both directions. Growth could be faster than estimated, or even accelerate, and the index used could underestimate the real impact of progress. Conversely, other factors than AI may have played a role, the tests relate to well-defined tasks, and the pace of Anthropic before the end of 2025 was perhaps abnormally low.

Transparency limited by redactions

The most discussed point of the report concerns what is not included in it. Anthropic had a unilateral redaction right and removed the majority of elements that METR considered most informative: detailed estimates, bottleneck analysis, algorithmic progress measurements and forecasts. The company explains that it wants to remove confidential information that could accelerate other AI developers.

METR recognizes that its report does not achieve the goal of transparency on the evidence, and that the public must rely on its judgment far more than would be desirable. The organization also said Anthropic canceled a planned audit of the distribution of its computing power between capacity development and security, wanting to focus the assessment on risks for the moment. METR did not have access to all employees or unpublished models, and some requests were refused. The full report, of around a hundred pages, was submitted on October 22, before a public version finalized on November 3.

Why this report matters

The idea of ​​an AI that accelerates its own research is at the heart of the concerns of security specialists: if each generation of models helps to design the next one more quickly, the time available to assess the risks is reduced. The METR report provides for the first time an external measurement of this phenomenon in a large laboratory.

It confirms that AI research is going faster, mainly because researchers are working with AI, without it yet carrying out entire sections of research on its own. But it also shows the limits of the exercise: without full access or publication of the evidence, the independent evaluation remains partial.

Questions that remain open

Several issues go beyond the sole case of Anthropic. Are other laboratories experiencing the same acceleration? Should we impose independent evaluations with wider access and public evidence? And how can we adapt AI security rules to a pace of progress that could accelerate further? Future evaluations, carried out with wider access, would make it possible to confirm or correct these first estimates. The response from laboratories and regulators will tell whether this type of external control can keep up.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles