Anthropic released Claude Haiku 5.5 on 7 October as a small model for tasks that run frequently and do not always require a flagship system: document retrieval, database questions, summarisation, compaction and tightly scoped coding subtasks. The API model identifier is claude-haiku-5-5, and this generation adds adjustable effort, letting developers explicitly trade response cost against deeper reasoning.
Anthropic estimates that Haiku 5.5 costs about 75% less to operate than Haiku 4.5 on average. It also cut Claude Sonnet 5.5 cache-read pricing in half, claiming about a 20% reduction in cost for typical agentic workloads. Those are vendor-reported workload averages rather than a promise that every application will see the same saving.
Why the subagent economics matter
The cost difference becomes material when one lead agent delegates dozens of classification, retrieval or context-reduction steps. Using a smaller model for those steps can reserve more expensive reasoning for tasks that genuinely need it. Anthropic's benchmark figures suggest substantial improvements over Haiku 4.5, but cross-vendor comparisons depend on test settings and have not been independently reproduced by Wire.
GitHub separately made Haiku 5.5 available to eligible Copilot customers, widening access beyond direct API users. Teams moving high-volume work onto the new model still need to test error rates and latency on their own workflows; a cheaper per-request price says little about the cost of retries or escalations.