The argument from a persistent slope change in AECI and leading indicators
The argument
A one-time jump predicts that models after Mythos Preview return to the old slope, while a sustained acceleration predicts they keep pulling further above it. Anthropic's benchmark scores have grown about twice as fast since October 2025, and newer internal models (later Model 2 snapshots and beyond) continue to widen the gap, so that a slope-change fit predicts held-out models better than a level-shift fit. Internal leading indicators also show elevated growth, which a single jump in benchmark capability would not by itself produce. Hence the trend is better described as sustained acceleration.
Does it work?
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported1 not separately assessed
Granting that newer models keep widening the gap over the old trend and that leading indicators show elevated growth, a sustained slope change becomes much more likely than a single level shift, though a short post-break window and a few data points leave room for a jump followed by noise. The argument rests on the roughly doubled benchmark growth rate, currently judged supported, and on the leading-indicator acceleration, not yet assessed and largely redacted.
Premises
- SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →Anthropic's models have improved on capability benchmarks about twice as fast since October 2025 as from March 2024 to October 2025.
- UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →Ratio of Anthropic's AI capability progress indicated by leading indicators since early-to-mid 2025 to its pre-acceleration rate
Where it is made
Passages where sources make, criticize or mention this argument. They record where it appears, not how strong it is, and none has precedence over another.
- Review of the "Risks from automated R&D" section in Anthropic's Risk Report (August 2026)makes itleaves a premise unstatedThe passage relies on a premise the inference needs without saying it.constitution →
At the time, the AECI data was only mildly in favor of a jump, and a continued acceleration could not be ruled out. If we reuse the data that was used in the Risk Report, the estimated acceleration is around 1.5x. After the publication of the Risk Report, Anthropic added data from more benchmarks, which estimates an acceleration of 1.9x using the same set of models.