METR on Anthropic: What the R&D Acceleration Report Actually Shows

A high-access look inside Anthropic's lab

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development.

METR is the nonprofit that tries to measure, in a scientific way, whether AI could cause catastrophic harm. Its staff spent August 12 through October 29, 2026 inside Anthropic on a high-access evaluation, then published the risk assessment on November 5, 2026. The question was whether AI speeding further development could make extreme, dangerous capabilities show up suddenly.

Anthropic opened internal research metrics to a degree METR describes as unprecedented. That included pretraining, reinforcement learning, and training compute broken out by domain. You are not getting a model release or an API. You are getting a written assessment of how fast the lab's own research stack may be compounding.

Risk conclusions you can read, evidence you cannot

Aimed at removing confidential information that could accelerate other AI developers.

METR was free to share its highest-level conclusions about risk. Anthropic kept redaction rights over non-public information and removed a majority of the evidence METR found most informative for those conclusions. Anthropic's stated aim was stripping confidential material that could speed other AI developers. Under METR's transparency approach, METR could disclose that those rights were used and whether findings depended significantly on redacted information.

Some redactions are hard to describe without exposing the sensitive material METR was asked to withhold. The report includes a table of important evidence it could and could not discuss, and it says every redaction that mattered for the conclusions is covered there or in in-line descriptions. Much of the withheld evidence had more than one plausible reading. METR and Anthropic researchers both remain substantially uncertain about how to turn it into one picture. METR is explicit that the report does not achieve transparency about the evidence, so you have to lean on METR's judgment more than is ideal if you want to check the claim against other sources yourself.

About 1.9 times the old pace, with a company reply still ahead

On the report's accounting, the capability-improvement measure AECI has run at about 1.9 times the historical pace since October 2025. That is the concrete speedup figure attached to this assessment, closer to a doubling than to a small bump, and it is why the evaluation is being read as more than a process note. Anthropic co-founder Sam McCandlish responded that the company will publish its own progress analysis. When that analysis will appear, and what it will contain, is not confirmed in the material released with METR's report.

The write-up runs from an executive summary through subjective risk conclusions, key findings, process details, and limitations. Appendices cover methodology, AECI benchmark scores, evidence of AI speedup, interview quotes on acceleration, adversarial robustness, process-transparency disclosures, and an AEF-1 checklist. Treat the pace figure as METR's synthesis under heavy redaction and shared uncertainty, not as a number you can recompute from a public dataset.

• Read the pace in context

  AECI since October 2025 is described as about 1.9 times the historical rate. The report does not hand you the unredacted series behind that multiple.

• Expect a second account

  Sam McCandlish said Anthropic will issue its own progress analysis. Timing and scope are still open.

• Use the appendices as the map

  Methodology, AECI scores, speedup evidence, adversarial robustness, process-transparency disclosures, and the AEF-1 checklist are where the report shows its work within the redaction limits.