Anthropic launches Claude Haiku 5.5 family with five models spanning wide performance and price range

Anthropic has unveiled Claude Haiku 5.5, a series of five models differentiated by intelligence, speed and pricing to let users match a version to the priority of each task. The gap between the cheapest and most expensive variant reaches 8.7× in cost per task, and the spread is designed to cover scenarios from immediate low-cost responses to maximum reasoning depth.

The most capable version, Haiku 5.5 (Max), scores 43 on Anthropic’s internal intelligence benchmark and delivers an output rate of 243 tokens per second. At the other end, Haiku 5.5 (Low) offers a first-token latency of 13.58 seconds and a cost of $0.02 per task. Three intermediate versions fill the spectrum, each targeting a different balance of depth, speed and price.

Anthropic defines four distinct comparison categories. Non-reasoning models are measured only against their peers; reasoning models are measured against both groups. Open-weight models are compared only within the same size class — Tiny (up to 4 billion parameters), Small (4–40 billion), Medium (40–150 billion) and Large (above 150 billion). Proprietary models are measured against both proprietary and open models in the same price band, using a weighted 3:1 input-to-output ratio across three tiers: below $0.15 per million tokens, $0.15–1 per million, and above $1 per million.

Four separate indices are reported, each aggregating 6-7 evaluations: AA-Omniscience, GDPval-AA v2.1, AA-Briefcase v1.1, Humanity's Last Exam, AutomationBench-AA, AA-LCR v1.1, GDP.pdf and MLCR-AA. The weighting differs across indices so that each emphasizes a different facet — general knowledge, coding, planning, or handling complex automation tasks. The company has not published full numerical results for every version individually, only the peak score of 43 and the latency and cost extremes.