Anthropic has released Claude Haiku 5.5, the newest version of its small, low-cost AI model, and its headline pricing undercuts its predecessor dramatically. The bigger question is whether that sticker price reflects what developers will actually spend.
Like Haiku 4.5 before it, the new model is built for high-volume work such as summarization and classification, where speed and cost matter more than raw capability. Anthropic calls it its fastest, cheapest and most capable small model yet. Pricing is split into two tiers: prompts of up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens, rising to $0.50 and $2.50 above that threshold. Haiku 4.5 launched at $1 and $5, so the lower tier is a tenfold cut on paper. Anthropic’s own estimate is more conservative, putting average running costs about 75 percent below the previous model.
That pricing matches OpenAI’s GPT-6 Luna, the obvious rival in this segment, and Anthropic’s benchmark figures show Haiku 5.5 ahead of Luna across all six tests it published. The biggest gains are in computer use and agentic coding. On OSWorld 2.1, Haiku 5.5 scored 72.4 percent against Luna’s 48.9 percent and Haiku 4.5’s 15.7 percent. On Terminal-Bench 4.0, where the older Haiku scored zero, the new model reached 39.2 percent, compared with 16.4 percent for Luna. These are company-supplied numbers, though, and that matters.
Independent testing from Artificial Analysis adds an important caveat. At maximum settings, Haiku 5.5 scored 43 on its Intelligence Index against Luna’s 38, but it used more than three times as many output and reasoning tokens per task. With identical per-token rates, that put its average cost per task at $0.21, three times Luna’s $0.07. It also generates tokens nearly twice as fast, yet still took longer on average to finish each task because it had so much more to produce.
In practice, per-token pricing is only half the calculation. For teams running agents at scale, how much a model “thinks” before answering can matter as much as the price on the rate card.
Anthropic also says Haiku 5.5 shows fewer signs of misaligned behavior and is less willing to help with potentially harmful requests than Haiku 4.5. Alongside the launch, Claude Max subscribers will receive monthly API credits of $100 or $200 depending on their plan, while Team plans get up to $500 in pooled credits.
Haiku 5.5 is available now on the Claude Platform, AWS, Google Cloud and Microsoft Azure. The independent figures cover only maximum settings, so it is still unclear whether lower effort levels can keep the intelligence advantage while closing the cost gap with Luna.
