Haiku 5.5 makes small AI jobs cheaper, with a migration catch

A creator’s automation often needs a small job done hundreds of times: sort a research note, condense a description or extract a field. Anthropic’s October 7, 2026 Haiku 5.5 launch targets that kind of bounded work. It also reduces Sonnet 5.5 cache-read pricing. Launch and availability.

The interesting decision is which step deserves the smaller model. Keep an editor or a more capable setup responsible for the final argument; consider a cheaper model for repetitive preparation with a clear acceptance check. That is our suggested workflow, not a performance result.

Price the prompt you actually send

Anthropic lists these API rates in USD per million tokens, before separate cache charges. Current pricing documentation.

Sonnet 5.5 cache reads fall from $0.20 to $0.10 per million tokens. That saving depends on cache use; it does not halve every Sonnet bill. The launch names AWS, Google Cloud and Azure availability alongside the Claude Platform model claude-haiku-5-5. Check your provider’s billing and access separately from a Claude app subscription. Pricing table and related change.

The migration needs more than a new name

Old token counts can mislead a budget. The migration guide says the newer tokenizer produces approximately 30% more input tokens for the same text, with the increase depending on content. Recount representative prompts using the new model. A previous output limit may also truncate a response. Token-count changes.

Thinking configuration changes too. Haiku 5.5 uses adaptive thinking by default; an old enabled setting with budget_tokens returns a 400 error. The guide also calls for removing temperature, top_p and top_k, and reading response blocks by type rather than assuming the first block is the answer. These are selected checks, not the entire migration. Request and response checklist.

Next action: choose ten familiar inputs and define what a usable result must preserve before changing a production default. For a research summary, check names, dates and source links, then compare the total cost including retries and corrections. Our AI-output evaluation guide helps set that standard; the Claude Code workflow guide covers repeatable checks for coding work.