Claude Opus 4.6 to Claude Opus 5.5: the list price fell 20%, the cost of a finished answer rose 38% to 91%

The first battery was 60 ordinary requests, and its result did not hold in size on a different mix. So we added 120 requests in four task types (30 each of code, summarising a real article passage, a 400-word explanation and short reasoning), answered by the same 9 models with a 4,000-token cap and no answer cut off. These are constructed requests, task templates crossed with topic lists, not real user prompts.

Opus 4.6 to 5.5 rose +38% here against +91% on the ordinary requests, so the direction held and the size did not. Sonnet 4.6 to 5.5 changed -16% overall, and the sign depends on the task: -32% for code, -34% for short reasoning and +12% for long-form writing.

Percentage change in cost per answer at list price, battery 2. The 95% interval is a paired bootstrap over the 120 requests; per-task cells hold 30 requests each and are wider.

Mean input and output tokens per answer over the 60 requests, the provider-stated list price, and the cost per answer in each run.

Battery 1 is 60 ordinary requests, the frozen battery of the Decay Index. Battery 2 is 120 constructed task-typed requests (battery2.jsonl). Prices are the provider-stated list prices archived by AI Price Watch. Full method: METHOD.md.

Why every headline is a range. A figure from one design is a figure about that design. The headline spans every usable design in data.json (the 60 ordinary requests at a 4,000-token cap, and the 120 task-typed requests), and the first capture at a 1,200-token cap is shown but not used because it cut off up to a third of some models' answers. A design counts as usable when no model in the step had more than 5% of its answers cut off; that line was set after we saw the counts, so it is a disclosed choice, not a pre-declared one. The page generator refuses to print a range that differs from the one derived in the data file.

Closure checks. The three-cause split rebuilds the provider-reported change in output tokens to within 3.9 points on every step, and the residuals are in the data file. answers.csv holds all 2,160 answers, as counted. Thinking tokens never exceed billed output tokens.

• Battery 2 is 120 constructed requests (30 each of code, summarising, long-form explanation and short reasoning), not real user prompts: task templates crossed with topic lists, with real article paragraphs as the passages to summarise. It was added because the first battery of 60 was thin and its result did not hold in size on a different mix.
• Cost per answer is the measured mean input and output tokens priced at list price. It is not a bill: caching, batch discounts, reasoning tokens and plan limits are ignored.
• It measures what answers cost, not how good they are. A longer answer can be a better answer; this dataset does not say.
• One battery of 60 ordinary requests, one capture per model per run, Claude models only. Run 1 is the frozen Decay Index capture (1,200-token cap, so some long answers were cut off); run 2 raised the cap to 4,000 and is the headline.
• Run 2 was captured before this method note was written, so the method was not committed in advance and this is not a pre-registered study.
• Temperature was 0 where the model accepted it and the provider default otherwise; it is recorded per answer. There is no clean run-to-run noise estimate because every model was cut off at least once in run 1; uncertainty is the paired bootstrap interval over prompts.
• Hidden thinking tokens are billed inside output_tokens. Their share was measured on a separate recapture of the 5.x models only; for the three oldest models the API reports no thinking count and thinking is taken as 0 because no thinking block was returned. The cost figures use the provider-reported output tokens, so they include thinking.
• Claude Haiku 5.5 has no list price in the archived tables, so it has token counts but no cost per answer; no price was guessed.
• Only input and output tokens of this battery are priced. The ratio between input and output depends on the task; short prompts with long answers (as here) make output price dominate.

Related work. Cost per task rather than per token is also published by Artificial Analysis, and the Opus 4.7 tokenizer effect on real traffic has been measured by OpenRouter and others. This dataset does not replace them. It adds the rows (every answer's tokens and cost, openly licensed), intervals, and a split of the change into tokenizer, visible writing and hidden thinking on a fixed battery.

No provider was asked to comment before publication. A provider or reader who finds a wrong figure can tell us; we correct it and date the correction.