Research group METR: Anthropic is probably improving its AI models twice as fast as before

Is AI already helping to make AI better faster? It is one of the biggest questions on the road to the singularity, because then progress can feed on itself. The non-profit organisation METR, which evaluates AI models, has now published an estimate for one company.

A small METR team spent time inside Anthropic, the company behind the AI assistant Claude, from 12 August to 29 October. Most METR staff who worked on it now think that Anthropic is improving its models on clearly defined tasks about twice as fast as between March 2024 and October 2025. In one year, that means about as much progress as was previously made in two years.

The uncertainty is large. It would not surprise the researchers if the pace were closer to 1.3 times or 3 times as high, and the team itself did not agree on everything.

Where the figure comes from

The main evidence is an internal yardstick at Anthropic, the AECI. It is the company's own version of an index from Epoch AI that tracks what AI models can do. On that yardstick, according to METR, growth since October 2025 has been 1.9 times as fast as before. It probably lies between 1.5 and 2.8 times.

Tests on a yardstick like this are often easy to check automatically. Real work is messier, and METR does not yet know how the acceleration plays out there. Also, before the end of 2025, Anthropic was growing more slowly on this yardstick than the leaders in the industry, which may explain part of the acceleration.

AI as a help in its own research

METR obtained convincing evidence that the acceleration does not come from Anthropic putting more computing power into its training more quickly. That evidence is not public. The team's best estimate is that the majority of the current acceleration, perhaps about 80 per cent of it, comes from the heavy use of AI in research at Anthropic. Among Anthropic staff, too, there was no agreement on this. For the acceleration at the end of 2025, other causes probably played a large role. METR looked at new staff, data and new algorithms, among other things.

AI works mainly as a tool here. On its own, it does not yet make a large contribution to the most important parts of training a model. Researchers did see clear signs that this could change quickly.

In polls on the internal chat, staff estimated that the newest model made them 12 times as fast in their work in October. METR calls this very loose evidence, because people gave it little thought. A personal figure like this also says something different from the pace of the whole company: in interviews, estimates for that were around 2 times.

No brake in sight, but a warning

METR saw no evidence of major bottlenecks in computing power, data or algorithms that would slow growth over the next twelve months. The team is not certain of this: its view changed several times during the assessment and it may have missed things. METR also sees a serious risk that progress speeds up even further in the coming year.

Much of the evidence stays secret

Anthropic gave METR unprecedented access to internal data, but was also allowed to remove non-public information from the report. This happened with the majority of the evidence that METR considered most important. Anthropic wanted to protect confidential information that could make other AI companies faster. As a result, readers have to rely on METR's judgement more than is ideal, the organisation writes itself.

Partway through, Anthropic also stopped a planned check of how much computing power it spends on safety and how much on better models. The company wanted to focus the assessment on risks for now and wants to return to that check later. There had already been criticism of the financial ties between METR and Anthropic. For this assessment, METR accepted no payment, but did get free use of the AI models.

What does this mean for the road to the singularity?

AI that helps to make better AI is the core of the idea behind the singularity. This report gives a concrete estimate from the inside, for one company and with a large margin of error. The next step is for such measurements to be made more often, at more companies and with more public evidence, so that everyone can follow how fast it is really going.

What is the technological singularity? Read the explanation →