This is a special post for quick takes by nostalgebraist. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
If you look at frontier ECI over time, there was a clear trend break in mid-to-late 2024 around the introduction of reasoning models. The slope increased by ~2x around this point[1], while being roughly constant before and after it.
If you then break things down by each lab's individual frontier, you'll see that the overall industry frontier's acceleration was really a manifestation of a more specific phenomenon: an acceleration for OpenAI models. OpenAI's acceleration allowed them to overtake others and dominate the frontier for a while after the trend break, which made "the frontier trend" and "the OpenAI trend" approximately the same thing over that interval.
Most notably, no such speedup in 2024 is visible for Anthropic models.
Anthropic's frontier did experience its own trend break, but only later on, in late 2025. This is visible in ordinary ECI, you don't need AECI or internal models to see it[2]. As far as we can see from the sparse/noisy data, it is similar to the trend break that OpenAI experienced in 2024: sudden ~2x slope increase, roughly linear before and after that.
Since Anthropic's pre-break slope had been a bit higher than OpenAI's pre-break slope, the same holds true for the two post-break slopes, and so Anthropic is now moving faster than OpenAI. That's allowed them to catch up, and now they've "become the frontier" in the same way that OpenAI "was the frontier" before.
I remember people freaking out (understandably) about the 2024 trend break once it became clear it had happened. And now, people are freaking out about the separate 2025 trend break, attributing it to coding agent assistance with AI development (an explanation no one ever proposed for its earlier sibling, since agents back then were just not that good yet).
But something doesn't add up about this... it's not as though "the frontier" sped up twice, and is now 4x as fast as the original trend! Instead, each of these two labs experienced a single ~2x speedup, at different points in time.
Why shouldn't I conclude that the Anthropic trend break noted in the METR report is just "the large-scale reasoning RL speedup," arriving later for Anthropic than it did for OpenAI?[3]
The prevailing narrative seems to be, instead, that the 2024 speedup was due to the introduction of reasoning RL, and the 2025 speedup was due to internal use of coding agents. But that story has to explain both...
...whereas the "it's all reasoning RL" story only has to explain one a-priori-surprising observation -- why the single mechanism for the speedup arrived at one time for one lab, and at another time for the other. Unlike the coding-agent story, this one does not struggle with the question "why did the 2025 speedup only happen to Anthropic and not OpenAI?", since on this story it had already arrived for OpenAI, in 2024.
(The data is so sparse and noisy that I feel uncomfortable reading this much into it, but that's a problem for any analysis of this kind. Not to mention the still-outstanding questions about "well-specified tasks," about y-axis units, about whether a piecewise linear curve with a constant slope from Sonnet 4.5 to the present is really what "the better your coding agent is, the faster you go" would look like in practice if it were true, etc. But many others have made those points elsewhere, while the argument I mention in this note seems weirdly under-discussed to me.)
The linked post is technically doing something slightly different where it fits two lines, one for reasoning models and one for non-reasoning models, rather than fitting a piecewise linear curve to points from a single frontier. I don't think this distinction matters that much for the conclusion, though.
As has been said elsewhere, including internal models arguably favors a "trend break" narrative over a "one-time jump at Mythos" narrative -- but any way you slice it, Anthropic made more progress over the last year than they did in the previous one.
Of course Anthropic had reasoning models for a while before their trend break, but my anecdotal impression had been that they didn't reason as deeply or persistently as OpenAI models until sometime around Opus 4.6; my very first impression of Opus 4.6 was "WOW this thing writes really long chains of thought and seems REALLY keen on double-checking and triple-checking its work in every way it can think of before finishing." I remember this very clearly because it was such a striking difference. (I had swapped Opus 4.6 into a pipeline that had previously been using Sonnet 4.5 or Opus 4.5, so it was a naturally paired comparison.)
And of course -- whatever Anthropic was doing from late 2024 to late 2025 -- it apparently did not yield the benchmark score speedup that OpenAI got out of their own reasoning RL, even if "reasoning" was technically involved.