Claude Opus 5.5's cost claims point to efficiency gains, not just price cuts

Anthropic released Claude Opus 5.5 on September 22, positioning it as a frontier model for long-running coding agents and professional knowledge work. Its benchmark and pricing data show a pattern worth separating out for practitioners: a modest list-price cut paired with a much larger claimed cost saving, driven by efficiency rather than pricing alone.

On paper, the price change is incremental. Input tokens dropped 20 percent, from $5 to $4 per million, and output tokens fell the same share, from $25 to $20. Cache-read costs fell further, by 60 percent, to $0.20 per million tokens.

Coverage of the release nonetheless described savings well beyond that list-price cut. VentureBeat reported the model as 60 percent cheaper on API pricing relative to the competing Fable 5.1, while The Decoder cited Anthropic's estimate of roughly 40 percent lower total operating cost versus the prior Opus 5. The gap between a 20 percent list-price cut and a 40-to-60 percent total-cost claim indicates where the actual savings originate: fewer tokens and fewer steps to complete the same task. Anthropic's own announcement puts a third figure alongside those two: a 40 percent list-price cut netting out to roughly 40 percent in total cost savings — directionally the same story as VentureBeat's and The Decoder's numbers, with the spread between 40 and 60 percent reflecting how much a given workload leans on the token-efficiency gains versus the sticker-price cut alone.

That efficiency shows up in the benchmark data. Opus 5.5 scored 66.4 percent on Terminal-Bench 4.0, ahead of Fable 5.1's 55.8 percent and GPT-6 Astra's 57.9 percent, and posted a GDPval-AA v2.1 score of 1,846 Elo against Fable 5.1's 1,735.

"The model used among the fewest tokens and steps measured, while solving more terminal tasks than Opus 5." — GitHub chief product officer, quoted by VentureBeat

That description lines up with the field results Anthropic reported: a 680,000-line code migration completed in under a day, and a 200,000-line codebase audit finished in three hours, down from more than 20 hours on Opus 5. For practitioners who track cost per completed task rather than cost per token, that distinction is the more useful measure of the release — on the Artificial Analysis Intelligence Index, Opus 5.5 posted 58 points, the highest score recorded on that index to date.