Anthropic released Claude Haiku 5.5 on October 7 for high-volume routine tasks. Model documentation and Simon Willison's own API experiment corroborate availability. We do not treat the vendor's comparative benchmarks as independently verified superiority.

The practical bill is what makes this interesting: a cheap answer that needs two human corrections may not be cheap work. Savings depend on input length, output length and retries. The official price list distinguishes prompts up to 100,000 tokens from longer contexts; one universal discount claim would be misleading.

Imagine a small maintenance business. AI sorts incoming requests, prepares a short job history and flags missing information. A technician starts with the real problem rather than an entire inbox. This is our possible application scenario, not a demonstrated deployment result for this model.

The business would first need representative and tricky test cases, correction measurements and a clear handoff to a person. Automatic payments or safety interventions are poor starting points. A sensible first step is suggestion-only mode without permission to modify production data, with spending limits and personal-data protection.

Our optimistic editorial horizon is 1–4 weeks for a limited pilot in a team with good data and existing integration capabilities. Reliable operation for every customer could take considerably longer. Real progress might be quiet: less copying and searching without more invisible mistakes.