Auditors Say SI Is Now Speeding Up Its Own Creation. Claude’s Maker Did Two Years’ Work in One.

For years the big worry in Silicon Valley has been a simple one: what happens when SI gets good enough to help build the next, smarter SI? Each one speeds up the next, and the whole thing starts running faster than people can check it.

Last Thursday an outside group said that loop has started at one of the biggest labs in America.

What the auditors found

The group is METR, a nonprofit that tests SI systems for danger. From Aug. 12 to Oct. 29, a small METR team worked inside Anthropic, the company behind the Claude chatbot, on company laptops, with access to internal scores, unreleased models and staff interviews. They took no money for it. Their public report came out Nov. 5.

The headline number: since October 2025, Anthropic’s models have been improving about 1.9 times as fast as they did before. In METR’s words, that implies “Anthropic has made roughly two years of historical progress over the last year.”

The auditors said they were “confident” the speedup did not come from simply buying more computer power to train the models. Their best guess is that “perhaps ≈80%” of today’s speedup “is due to AI,” meaning Anthropic’s own SI doing research work alongside its people. They stressed that the estimate carries “substantial uncertainty,” and that the true pace could be anywhere from about 1.3 to 3 times the old rate.

What the people inside said

Anthropic staff vote in quick Slack polls on how much faster the newest model makes them. In October 2026 the average answer was about 12 times as productive, roughly double what they reported a few months earlier. METR called those polls “only very loose evidence,” but the trend kept climbing.

The interviews were blunter. “It is hard to see how I am not being 3x more productive,” one researcher told METR. “I feel like I have postdocs and grad students.” Another said the team is “now only bottlenecked on ideas & compute, no longer bottlenecked on code.” Not everyone agreed. One researcher guessed only “15-20% is due to AI” and said they “don’t believe in self-reported productivity increase.”

For now, people still steer. METR wrote that the SI “is not yet able to make highly significant and fully autonomous contributions to the most important parts” of building a model. But researchers described “signs of life,” meaning such contributions “could occur soon.”

Why it matters

Here is the line that should get Washington’s attention: METR said it found no clear roadblock that would stop things from speeding up further over the next 12 months, so there is “a significant risk that further capabilities progress could lead to even faster acceleration.”

And the public only gets part of the picture. Anthropic had the right to black out anything confidential, and METR said the company “redacted a majority of the evidence we found most informative.” Anthropic said those cuts were “aimed at removing confidential information that could accelerate other AI developers.” Anthropic also dropped a planned piece of the review that would have checked how much of its computing goes to safety versus building more power, saying it would revisit that later.

So the outside check happened, and it’s more access than any lab has given before. But on the biggest question in tech right now, how fast this is moving, Americans are mostly being asked to take the auditors’ word for it.

Read METR’s full public report here.