Given how widely varying the amount of tokens each model uses for a given task, "price-per-token" is essentially meaningless when doing this sort of comparison.

Artificial Analysis's "Cost to run" model (aka num_tokens_used * price_per_token) is much better, but even that is likely problematic since it's not clear whether running a bunch of benchmarks maps cleanly to real-world token use.

That graph seems odd. It looks like Gemini 3.5 Flash is not actually on the convex hull, and they forced the 'frontier' to bend inwards to include it