Comment by Dibes
4 days ago
I'm not sure what to make of this graph[0]. It shows medium as the most effective thinking mode by far for frontier code.
It's the only case that I saw going through the system card where more reasoning effort meaningfully negatively impacted the resulting eval. I know sometimes max efforts show a small dip, but this is substantial. I wonder why in the world that is?
I'm not sure about the answer here, but this can be caused by the scoring rubric used by given benchmarks. For instance, if a benchmark docks scores for running too many commands or using too much wall-clock time, higher efforts will get lower scores.
Pure speculation, but I've noticed drawbacks to the models on high effort. I interact mostly through prompts rather than agents so I sometimes see where their reasoning falls short. A model on high effort has longer output and can get hyperfocused on irrelevant details, maybe increasing the surface area for mistakes. I haven't used other effort levels extensively yet but I've supposed that medium may have more balance.
Very interesting. I mostly interact through prompts too but haven't changed the effort that much. I just assumed higher was better at the trade off of tokens so I have largely left it at the default.
It's "Cost per task", so perhaps it burns tokens too quickly, trying to "do a better job". Over-engineering :)
Last week it felt like Opus 4.8 was moving the Pro "usage" meter very quickly. Today, pre-announcement, Opus 4.8 Medium felt like there was less meter-use per minute. And post-announcement, Opus 5 Medium also feels more efficient, allowing more work in the 5-hour window.
Completely subjective, of course.
Just read something, my subjective view was correct, but it's not the model:
"Your weekly Claude Code limit is 50% higher through August 19"
apparently it got docked points for editing files out of scope
This must not be weighted very heavily on the benchmark because if it was, Opus would bomb every test (half kidding)
Do you have a source for this? That would explain it, but could be a bit of a concern on the general focus the model at higher thinking exhibits.
Maybe it's akin to the Ballmer peak: improved performance at a specific level of relaxation
I agree, very odd they did not comment on any theories for the degradation here. Dip and then rebound at max effort is pretty interesting too. Overthinking is bad, but you can overthink so much it starts to be better again?
It is probably from randomness. The benchmark tasks nowadays are so long that you can't really afford to run a large number of samples of them per model & effort combination
No it's a mean of 5 runs.
> We report FrontierCode’s overall score, a composite measure that grades each patch on blocking functional criteria (held-out unit tests) together with weighted code-quality rubric criteria, as mean@5.
They don't explain more in the system card, I guess higher effort levels could loose points on the code quality / scope / style / maintainability stuff?
If you're doing passes@1, especially for long-horizon agentic benchmarks, you might as well as not do the benchmark at all.