Haiku 5.5 cuts token prices 90% on prompts under 100k

On short prompts, input now costs $0.10 per million tokens against $1.00 on Haiku 4.5, and output costs $0.50 against $5.00. Anthropic puts the average saving at around 75%. Prompts over 100k get a smaller cut, and The News Stack reports that an updated tokenizer uses slightly more tokens per task.

The capability gap shows on Terminal-Bench 4.0, where Haiku 5.5 scores 39.2% and Haiku 4.5 scored 0.0%. Anthropic still recommends Sonnet 5.5 and Opus 5.5 for hard agentic coding. It aims Haiku 5.5, the first Haiku with an effort setting, at compaction, summaries and subagent work.

Anthropic also halved cache reads on Sonnet 5.5, from $0.20 to $0.10 per million tokens, and says that makes most agentic work around 20% cheaper. Max and Team subscribers get monthly API credits this week. Only the token count per task went up.