Claude Haiku 5.5 puts the cost of useful AI work under pressure.

Anthropic introduced Claude Haiku 5.5 as a lower-cost model aimed at common, high-volume AI workloads such as classification, summarization, document questions and delegated agent steps. The company estimates around 75% less cost per task on average compared with Haiku 4.5, drawing on its pricing and evaluations.

The phrase 'per task' is important, but the number should not be read as a universal discount. Prompt length, answer length, caching, retries, escalation and the share of difficult cases all change the economics.

Many useful applications are held back not because a frontier model cannot perform them, but because doing so thousands of times is too expensive. More capable small models can make mundane workflows—support triage, file processing and monitoring—economically feasible.

However, cheap calls are not cheap work if their errors require more human review. Real productivity depends on how many acceptable results a workflow produces, at what total cost and with what risk.

Cortist would compare Haiku 5.5 with a current baseline across a representative task mix, using identical quality thresholds. Measure accepted completions, retries, escalations, review time and overall spend.

A clever deployment may use the smaller model for routine cases and route uncertain or high-stakes work to stronger systems or people. The model does not need to win every task to improve the system.