Claude Opus 5.5 saves money on reasoning tasks, and writes only 11 percent faster than Opus 5

Anthropic priced Opus 5.5 as a cheaper, faster upgrade to Opus 5, and the price cut turned up in a real bill. The speed did not turn up in the same test.

Jessica Wachtel at The New Stack ran both models through three reasoning problems on 26 September 2026 and paid the Anthropic API for every token. She calls both from the API with identical prompts, adaptive thinking at default effort, and runs each problem once per model. Opus 5.5 does not let a caller turn thinking off, so both models get to think. Anthropic says the new model costs 40 percent less and runs 30 percent faster than Opus 5.

What the bill said

Opus 5.5 lists at $4 per million input tokens and $20 per million output tokens, down from $5 and $25. The price cut alone accounts for 20 percent of the saving Anthropic quotes; the rest has to come from the model using fewer tokens. On the logic grid puzzle both models answered correctly, and Opus 5.5 used 7,573 output tokens against 10,621 for Opus 5, for $0.16 against $0.27. That is 43 percent cheaper for the same right answer. On the stone game both again finished correctly, and Opus 5.5 used 28,740 tokens for $0.58 against 74,981 tokens and $1.88 on Opus 5, 69 percent cheaper. Across the whole run Wachtel paid $10.50.

What the stopwatch said

The token-per-second numbers came out closer than Anthropic's headline. Opus 5.5 wrote 103.4 tokens per second and Opus 5 wrote 93.1, so the new model is about 11 percent faster on writing speed. Its largest lead on any single problem was 19 percent. The 30 percent Anthropic quotes did not show up in this test.

The third problem, a counting task with answers 27, 1,695 and 159,019, exposed the failure mode a caller should plan for. Neither model finished it. With a 48,000-token output limit both spent the whole budget thinking and returned nothing. Wachtel raised the limit to 128,000 tokens and reran. Opus 5 used every token, thought for over 25 minutes and returned no answer, at a cost of $3.20. Opus 5.5 thought for 19 minutes, used 112,733 tokens, and the API ended the response with a "refusal" stop reason and no text. The prompt was benign, so the refusal reads as a mistake. Cheaper does not help when the answer never arrives.

What to do about it

The switch is worth making if a team already runs Opus 5. Both models scored the same on the three tests, so the ceiling on reasoning did not move; the cost of getting to a right answer did. Set a hard output limit and watch spending on hard problems, because either model can burn 20 minutes and return nothing on the wrong kind of question. For counting problems, hand the model a code execution tool instead of asking it to reason through every ordering, since a short program finds 159,019 in seconds where thinking does not find it in half an hour.

Anthropic's own benchmarks put Opus 5.5 ahead on coding, knowledge work and reasoning; those are not this test. What this test shows is that the vendor price claim held up on three real problems and the vendor speed claim did not.

Source

Claude Opus 5.5 vs. Opus 5 on reasoning tasks: Cheaper, faster, but not better by Jessica Wachtel, The New Stack, 26 September 2026.