AI effective cost per answer 2026: what a finished answer costs, not what a token costs

What a finished answer costs, not what a token costs: nine Claude models answering two batteries (60 ordinary requests and 120 requests across code, summarising, long-form writing and short reasoning), priced at the provider-stated list prices archived by AI Price Watch. Includes every answer's input tokens, output tokens, stop reason and cost, the raw captures, and a split of the change in output tokens into the tokenizer, visible writing and hidden thinking. Headline: from Claude Opus 4.6 to 4.7 the list price did not change and the cost of a finished answer rose 26% to 39%; from Opus 4.6 to 5.5 the list price fell 20% while the cost rose 38% to 91% depending on the task mix; for Sonnet 4.6 to 5.5 the sign depends on the task. Method, limits, intervals and tables: https://nesyona.com/research/ai-effective-cost-2026/ . Limits: Claude models only; list prices, not bills; battery 2 is constructed requests, not real user prompts; the method was written after the second capture, so this is not a pre-registered study. Related work includes Artificial Analysis cost-per-task data and OpenRouter's Opus 4.7 tokenizer analysis. Released under CC BY 4.0. Zenodo stores a flat file list, so folder paths are flattened with a double underscore.