Hacker News Reader

Comment by ianm218 | Hacker News Reader

Why do these benchmarks often show models running on Max i.e. Opus 5.5 (max) is on the frontier? I would think it is never really economical to run max

View on HN

By definition, the model with the best performance will be on the pareto frontier (regardless of cost).

No. Pareto frontier is cost performance tradeoff.

What jonas21 is stating is correct. If you have the absolute best performance, you are automatically on the Pareto frontier because you’re the most cost effective one (or the only model) on that performance tier.

Huh, I guess you're right, but I interpreted jonas21's comment in relation to the above comment, which was about all models running on Max (which implies this would be a comment relating to all points on the pareto curve).

It is true that the model with the overall best performance is on the Pareto frontier, but it is not necessarily true that a given model's best cost-performance tradeoff is at max reasoning. (IE you could have a model that is on the frontier at medium reasoning, but at max reasoning becomes more expensive and not sufficiently good enough to remain on the frontier.)

If you're the best then you'd be thr first regarding of cost since there is nothing to compare to for that performance.

As other have stated, your point is relevant only for the single best performance point. The question is for interior points of the frontier where the tradeoff actually exists. For instance why is GPT 6 Luna (Max)on the frontier rather than GPT 6 Luna (Medium)? It is not because GPT 6 Luna (Max) is the most performant model regardless of cost.

Why is this downvoted? From my understanding dominated solutions aren't on the paerto frontier e.g. in this case where argon and opus 5.5 are tied the lower cost argon should be the end of the curve.

From the graph it's a flat line, but looking at the point yields a negligible increase with no error bars (indeed if you look at fable 5.1 it's measured as lower than opus and argon). This just seems like bad data or blatant favoritism at worse.

Yikes.

Check out slopalytics. You can draw a pareto line there with max reasoning excluded.

Reply on news.ycombinator.com