An on-call developer gets a report full of observations, guesses and steps already attempted. A subagent can list the symptom, version and past attempts. Anthropic launched Claude Haiku 5.5 on October 7, 2026, for repetitive, high-volume work such as summarization and classification. Its effort setting adjusts how much work the model does. Anthropic also recommends it as a subagent. The launch scores reflect the vendor’s tests and settings.
Test with your own data
Anonymize 30 reports and manually label facts to preserve. Include missing versions, unconfirmed guesses and reports with multiple issues. With the same instructions, compare your current model and two Haiku effort levels. Check the category, facts omitted and facts invented. Measure time to complete the entire task, including any escalation or manual edit.
Making a decision
Set a limit for version errors and calculate cost per accepted report including rework. If correction takes longer than expected, start with reports whose fields are reliably filled. Rerun the same cases after model or price changes.