Artificial Analysis ships side-by-side model comparison tool

The new release comparison page from Artificial Analysis puts up to five model releases side by side, and it lists every variant of each one rather than a single entry per model. That matters because many current models run at several effort settings, which trade more thinking for higher cost and slower answers. In the snapshot collected for this item, comparing two releases expanded into ten variants, one per effort level, each with its own scores, prices and timings.

For each variant the page shows the company's Intelligence Index, a composite of ten evaluations, alongside the individual benchmark scores and industry indexes such as Finance and Accounting, Legal, Healthcare and Engineering. The cost side includes per-token prices for input, output and cached tokens, and also a cost per task, which the page defines as a weighted average cost per Intelligence Index task calculated from those token prices. Because it reflects how many tokens a variant actually uses, it can separate settings that share the same per-token price: in the collected snapshot, one model's highest and lowest effort settings had identical token prices but differed roughly tenfold in cost per task.

The announcement lists the same categories. The page is a live tool, so its figures will change as models and measurements are updated, and all of the indexes are Artificial Analysis's own measures. The announcement's video could not be extracted and was not reviewed.