METR: Anthropic Has Nearly Doubled Its Pace, and 80% of the Acceleration Came From AI Inside the Company Itself

Imagine your team is writing a compiler and suddenly realizes that the new version of the compiler builds the next one faster than you can check the results. This is exactly the scenario AI safety researchers fear, and it is exactly what METR was examining at Anthropic.

What the auditors found

From August 12 to October 29, 2026, the METR team (about 2.5 full-time staff) was given unusually deep access: benchmark results for non-public models, training metrics, employee surveys and roughly 15 interviews with researchers. The organization took no money from Anthropic, but it worked on company devices.

The main conclusion: since October 2025, Anthropic's model results have been growing about 1.9 times faster than in the period from March 2024 to October 2025. On the composite AECI metric, the gain rose from 15 to 29 points per year. In effect, in one year the company covered ground that previously took two, roughly like the leap from Claude 2.1 to Sonnet 4.5.

This is not explained by growth in training compute. By METR's estimate, about 80% of the acceleration comes from the use of AI in Anthropic's own research. One employee said: "Now we are limited only by ideas and compute, not by code." Experiments that used to take one or two days to set up have become almost free. In informal Slack polls, employees put the productivity gain with the latest internal model at roughly 12x, but METR calls this "very weak evidence": the sample is unrepresentative and most likely inflated.

The threshold that may have been crossed in July

In its Responsible Scaling Policy, Anthropic stated that if the pace of progress doubles because of AI research automation, this is grounds for a warning. In its public August report, the company wrote that the acceleration was "less than two." But the internal unredacted version gives 1.97, and Anthropic itself considers it "plausible" that the threshold has been crossed. METR allows that this may have happened as early as July 15, and faults the public report for understating both the acceleration and the uncertainty.

For now AI works mainly as an assistant rather than an autonomous researcher, although there are "signs of life" of independent contributions. The auditors found no bottlenecks that would slow the pace over the next 12 months, so a "substantial risk" of further acceleration remains. METR, like Anthropic, rates the acute catastrophic risk as low.

Where the data is shaky

Uncertainty is large: the true multiplier could be anywhere from 1.3 to 3. If models from OpenAI, Google and xAI are included, the industry-wide figure falls to 1.4. It may be affected by a catch-up effect: in 2025 Anthropic was lagging behind. Opinions inside the company differ: one researcher puts AI's contribution at only 15–20% and does not trust colleagues' self-reports, while another sees new hires as the main factor.

The most painful part is the redaction. Anthropic had the right to unilaterally redact any non-public information, and it used that right extensively. Most of the material the auditors considered the most informative was removed: forecasts for future models, a breakdown by training steps, and bottleneck analysis. The company explains this as protecting information that could speed up competitors. As a result, the public has to trust METR's judgment "much more than we would like." One of the charts in the report is labeled in large letters "FAKE DATA."

In addition, Anthropic canceled part of the project: a check of how the company divides compute between capability development and safety research. Daniel Kokotajlo, author of the "AI 2027" scenario, sharply remarked that without this audit it is impossible to verify commitments to slow down.

What this means in practice

For developers and founders the takeaway is simple. If code is no longer the bottleneck at a leading lab, the model release cycle will shrink, and APIs and capabilities may change more often than your annual plans assume. Build an abstraction layer over the provider into your products and run regular regression tests on your own tasks. Product managers should remember that figures like "12x more productive" are self-reports, not measurements. And finally, an external audit in which the audited party holds censorship rights is so far more partial transparency than a guarantee.