Why Opus 5.5 cost per task barely moved despite price cuts

Claude Opus 5.5 at its maximum effort setting uses so many more tokens than Opus 5 that Anthropic's price cuts only bring a task's cost back to where it was. Artificial Analysis, which benchmarks AI models, puts the cost at $5.98 per task in its Intelligence Index for Opus 5.5 (max), against $5.86 for Opus 5 (max). The measure is what it costs to have a model complete one task in the firm's own evaluation suite, so it reflects both the price per token and how many tokens the model spends getting to an answer.

The firm splits the change into three steps. At Opus 5's prices, Opus 5.5's extra token usage alone would raise the cost per task by about 80%, to $10.51. The 20% cut to base prices brings that to $8.41. Lowering the price of cache reads to $0.20 then removes a further 29%, landing at $5.98. Cache reads are charges for input the model has already processed and stored, such as a long conversation or codebase it re-reads on each step. A price cut can remove no more than the charge it applies to, so cache reads must have made up at least 29% of the $8.41 cost before that last cut.

So the cheaper token prices that Artificial Analysis covered in its launch-day post on Opus 5.5 do not mean cheaper work at max effort: they pay for the model's heavier token use. These are the firm's figures for its own index at max effort only. Workloads that lean less on cached input, or that run the model at a lower effort setting, could see a different balance, and the post does not give those numbers.