Claude Haiku 5.5 arrives, breaking six things from Haiku 4.5

Claude Haiku 5.5 is the first Haiku model to adopt the same API shape as Opus 5.5 and Sonnet 5.5, which means it inherits the same migration pain those models introduced. Haiku 4.5 was the last generation where budget_tokens, temperature and prefill all worked together, so developers moving only the model ID will encounter more breakage than with other 5.5 models. The official migration guide, dated October 12, 2026, lists six changes that now return HTTP 400, while other behaviors change silently — thinking turns on by default, token counts rise roughly 30%, and thinking blocks from all turns stay in the input. A note says the author tested two of the six errors and the disabled-effort limit against the live API with short requests, confirming that failed calls are not billed.

Another silent change is a new refusal category covering cyber, frontier_llm, bio and general_harms, which returns HTTP 200 with stop_reason: "refusal" and no server-side fallback. On pricing, Haiku 5.5 costs more when a single request's input tokens, including cache reads and writes, exceed 100,000 — the first model with that threshold. Long-running agents should consider compaction or compare against Sonnet 5.5, and the "Haiku is 1/20" rule of thumb no longer holds. The author's own code only used Haiku 4.5 in one patrolling agent worker with no thinking or temperature, so a model ID change was enough there.