If we start at o1-mini (the first reasoning model) we get a much cleaner fit, and the slope change becomes 1.9x as well!
Using ECI (with the public release date) results in the slope change being only 1.4x, but mostly from Fable 5.2 taking much longer between internal and public deployment
I'm assuming Model 3 Early is the same as Fable 5.2, which they basically say in the system card
Using only models which have ECI scores (excluding internal models and Gemini 3 Deep Think), and using the internal release date for Fable 5.2 gives a 1.82x slope change for AECI and 1.67x for ECI.
Anyways on rereading METR's report I realize I missed the importance of "likely" here (in hindsight probably because influenced is so soft), which I take to mean >50% chance
The acceleration in late 2025 was likely influenced by non-AI factors. We investigated a variety of factors which could have contributed to the observed acceleration in late 2025, but details of many of these factors have been redacted.
This is totally consistent with nostalgebraist's claim that RL scaling was the reason for the steeper trend with Opus 4.6! METR only says (emphasis mine) "a majority (perhaps ≈80%) of the current(October 2026) acceleration is due to AI"
As an example Mythos 5 being above trend could've been 70% from RL scaling and from 30% AI R&D, while Mythos 5.1 was 60% from RL scaling, Mythos 5.2 was 40%, etc. until you get the 80% from R&D in October.
It seems weird that they'd cancel each other out so well, but I don't think it's weirder than Moore's law being so straight (despite many S-curves being involved on the inside at various points)