METR: Anthropic has doubled the pace of model development, largely thanks to AI

For years, talk of “AI building AI” was a speculative scenario. Now there is a report from an independent organization that had access to the internal data of one of the leading labs. Its conclusion is cautious but alarming: recursive self-improvement appears to have already begun, albeit in a mild form.

What exactly METR found

From August 12 to October 29, 2026, METR staff worked inside Anthropic with expanded access. They were shown pretraining and RL metrics, compute allocations by workstream, internal model benchmark results, analysis of AI-driven acceleration and forecasts for a future model, and they interviewed employees.

The headline figure: since October 2025, Anthropic's model scores on its internal AECI benchmark—its version of the Epoch Capabilities Index—have been improving 1.9 times faster than before. The 95% confidence interval ranges from 1.5x to 2.8x. Put simply, over the past year the company has covered roughly two years' worth of progress at its previous pace.

Could Anthropic simply have spent more on hardware? No. METR estimates that the acceleration is not linked to increased training compute. In interviews and surveys, employees report multiple-fold gains from AI tools, which is consistent with the doubling in pace. Self-reports are difficult to interpret, however: people tend to overestimate their own productivity.

Why this is more than “just another estimate”

For now, AI mainly assists people rather than working entirely on its own. But according to the authors, there are already signs that its autonomous contribution will soon increase. If so, the acceleration could reinforce itself: a better model means faster research, which means an even better next model.

Most concerning, METR found no obvious brakes over the next 12 months—not in algorithmic ideas, training and experiment compute, or data. At the same time, the authors candidly acknowledge that their view of potential constraints changed several times as they received new information, so confidence on this point is low.

The reaction within the community is revealing. Ryan Greenblatt, who works at METR, acknowledged that he had expected only a 1.15–1.3x acceleration in 2026. Ajeya Cotra believes that, across the company as a whole, R&D may be moving 4–5 times faster. Daniel Kokotajlo of the AI Futures Project says the report confirms his concerns about short timelines to superintelligence.

Where the report is weakest

The first problem is that Anthropic removed some of the data underpinning the report's main risk conclusions. Greenblatt explicitly writes that conclusions about precisely where the acceleration comes from now depend heavily on trusting the authors. The public version should be read with that caveat in mind.

The second issue is the index itself. AECI measures verifiable tasks, which are not the same as real research progress. In addition, the slow pace from 2023 to 2025 may have been anomalously low, making the current acceleration appear more dramatic than it is. On the other hand, the index has sometimes understated real-world practical impact, so the error could run in either direction.

What practitioners should do about it

For developers and product teams, the takeaway is practical. Plan products on the assumption that models will be updated significantly more often than in previous years, making dependence on a single API version and its behavior a risk. Maintain a quality-evaluation layer and the ability to switch quickly between models.

For founders, if the acceleration is real, the cost and capabilities of AI tools will change faster than budgets anticipate. Internal processes should be measured the way Anthropic measures them: not through a team's impressions, but with clear metrics taken before and after AI deployment.

Most importantly, track transparency as well as the numbers. The dispute over the report's redactions will show whether labs can provide the outside world with enough data for such assessments to be independently verified.