Anthropic ships Haiku 5.5 with 75% lower running cost

Anthropic's Claude line has seen a steady cadence of updates to higher-end models such as Fable, Opus and Sonnet, while Haiku — the smallest and fastest tier — had gone roughly a year without a refresh. The arrival of Haiku 5.5 therefore fills a gap at the cheap end of the lineup, where customers doing repetitive, high-volume work had been running an older generation.

The model is designed for tasks where speed and cost matter more than peak reasoning: summarization, data compression, database queries and classification. In coding, it can act as a sub-agent alongside Opus 5.5 or Sonnet 5.5, handling cheaper steps so the expensive models are reserved for harder work. Its benchmark scores show a large jump over Haiku 4.5 and even a lead over GPT-6 Luna, though they remain below Sonnet 5.5.

Haiku 5.5 also introduces effort-level settings to the Haiku line for the first time, letting users trade cost against capability. The real test is whether the price advantage holds for large production workloads, where token volumes can push requests past the 100,000-token threshold — a point at which the cost per token rises, though it is still half of Haiku 4.5. Separately, Sonnet 5.5 now has its cache-read fee cut in half, which Anthropic says makes it about 20% cheaper to run for most agent work.