On November 5, 2026, METR published an assessment of AI development at Anthropic. It judges progress on well-specified tasks to be roughly twice as fast as of October 2026 compared with March 2024–October 2025, with substantial uncertainty of around 1.3–3 times. Anthropic redacted much of the most informative evidence. That limits independent public scrutiny of the conclusion.
How acceleration can be investigated
METR’s previously published October 14 methodology separates four questions: how quickly capabilities improve, how much of the change comes from AI, which constraints could slow future development, and whether the method would detect much stronger acceleration at all. It proposes combining benchmarks, intermediate indicators and researcher interviews. Labor, compute, data and algorithms are considered separately.
This distinction matters to readers. A higher model score does not automatically explain its cause. When the team, training process and tools change together, their contribution cannot be reduced to one number merely because trends coincide. This is our editorial interpretation of the causality problem, rather than a new quantitative conclusion about Anthropic.
What a capability index measures
Epoch AI describes the Epoch Capabilities Index as a common scale combining multiple AI benchmarks. It aims to support model comparisons across periods in which individual tests become saturated. The value itself is abstract; the documentation shows how it relates to expected scores on particular benchmarks. Such an index makes tracking easier, while the choice of tasks remains part of what the measurement means.
A real product may raise a different question: does its agent complete a particular job correctly, with how many retries and how much human supervision? A capability index and production efficiency are not interchangeable units. Even when both improve, their relationship needs to be investigated on the relevant work. A headline about research progress should therefore retain its measurement context when discussed by teams building ordinary applications.
External access and public scrutiny
In its separate policy on working with AI companies, METR says it should disclose access limitations, redactions and conflicts of interest. It distinguishes transparency about the process from transparency about evidence itself. The organization says it does not accept payment or donations from frontier AI companies it works with, but depends on their voluntary cooperation and permission to publish information.
These conditions help readers assess a report without guaranteeing that every judgment is correct. An external assessor may receive more internal data than an ordinary observer. The public does not necessarily gain the same ability to repeat the analysis. Those two levels of access need to remain distinct.
Disclosure also changes how a disagreement should be interpreted. A reader who cannot see a crucial input may reasonably remain uncertain without dismissing all of an assessor’s work. Conversely, receiving internal access does not make an assessment equivalent to a fully reproducible public experiment. The useful question is which conclusions can be checked directly and which depend on the evaluator’s judgment.
Editorial interpretation: what to watch next
A useful follow-up would include comparable future assessments, clearly described tests and observation of whether the trend persists. A methodological change should explain its effect on comparisons over time. New public evidence should also permit reconsideration of the original conclusion.
The report matters to the debate about AI assisting its own development. It does not justify promising twice the productivity for every programmer or declaring a particular future inevitable. A responsible reading keeps the assessment, its uncertainty and the boundaries of visible evidence together.
