Opus 5.5 matched Opus 5's puzzle answers for up to 69 percent less at its lower default effort

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens through Anthropic's API, down from $5 and $25 for Opus 5.

Anthropic claimed Opus 5.5 costs 40% less than Opus 5 on typical workloads and generates output more than 30% faster.

The New Stack called both models through the Anthropic API with identical prompts, adaptive thinking at the default effort level (Opus 5.5 does not allow thinking to be turned off), and ran each of three reasoning problems once per model.

With the limit raised to 128,000 tokens, Opus 5 used every token, took over 25 minutes, and stopped with no answer, costing $3.20. Opus 5.5 ran for 19 minutes and used 112,733 tokens, but the API ended the response with a 'refusal' stop reason and no text.

Anthropic's prompting guide says effort is the main control for how much Opus 5.5 thinks, that thinking is always on, that medium is the default on Opus 5.5 while Opus 5 defaults to high, and that developers should set effort explicitly and test several levels against their own evals rather than carrying over the Opus 5 setting.

In Anthropic's testing, Opus 5.5 at its default medium effort matched or beat Opus 5 at high effort on multistep agentic coding tasks, in fewer steps and with fewer tokens.

Thinking tokens are billed as output tokens.

The price cut alone accounts for a 20% saving; the rest of the claimed 40% saving has to come from the model using fewer tokens.

On the logic grid both models got all 28 cells right. Opus 5.5 took 65 seconds and 7,573 output tokens for $0.16; Opus 5 took 108 seconds and 10,621 output tokens for $0.27. Opus 5.5 was 43% cheaper.

On the stone game both models answered all three parts correctly. Opus 5.5 finished in 215 seconds with 28,740 output tokens, costing $0.58; Opus 5 took 624 seconds and 74,981 output tokens, costing $1.88. Opus 5.5 used 62% fewer tokens and cost 69% less.

On the constrained-orderings problem with a 48,000-token output limit, both models spent the whole budget thinking and never replied. Opus 5.5 used 489 seconds and $0.96; Opus 5 used 553 seconds and $1.20.

The correct counts for the orderings problem are 27, 1,695 and 159,019; the third is very hard to reach by reasoning alone, a program that checks every possible ordering can find it, and neither model could run code in the test.

At Opus 5.5's list output price, 112,733 tokens come to about $2.25.

The New Stack's test compared Opus 5.5 at medium effort with Opus 5 at high effort.

Opus 5.5 used about 29% fewer output tokens than Opus 5 on the logic grid.

On the 48,000-token capped run, Opus 5.5 cost exactly 20% less than Opus 5, equal to the list-price cut, because both billed the full 48,000 output tokens.

The New Stack's writer said the prompt asks the model to count job orderings and contains nothing sensitive, so the refusal was most likely a mistake by Anthropic's safety filter.

Anthropic's Opus 5.5 prompting guide includes a 'Safeguard refusals' section for requests that return stop_reason: "refusal".