The price of intelligence
Most model comparisons rank on quality and list price separately, which quietly hides the only question that decides a build: what does the next point of intelligence actually cost? Here both live on one chart, with the efficient frontier drawn, so a model that is beaten on price and quality has nowhere to hide.
Moving from DeepSeek V4 Flash to Claude Opus 5.5 costs 41x more per million tokens and buys 66% more measured intelligence. Of the 32 models that publish both a score and a price, 7 are worth considering. The rest are beaten outright by something cheaper and cleverer.
Hover a point for its pricing. The line traces the best score available at each price.
Scores are the Artificial Analysis Intelligence Index, an independent evaluator, not Timace’s own judgment. Prices are the standard published API rate and exclude batch discounts, prompt caching and committed-use agreements, all of which move real spend a long way. Only models publishing both a score and a price appear, so this is a view of the priced market rather than of every model in the catalog. Free and self-hosted models are excluded: their cost is real but it is not per-token.