Anthropic has released Claude Haiku 5.5, describing it as its fastest and most affordable small model. The launch pairs benchmark gains over Haiku 4.5 with lower token prices, aimed at teams running high-volume tasks such as summarization, database queries, classification, and live customer support.
Lower token prices come with a tokenizer caveat
Anthropic says Haiku 5.5 costs about 75 percent less on average than Haiku 4.5. For prompts up to 100,000 tokens, prices fall by as much as 90 percent. The company says requests of that size made up roughly 90 percent of previous Haiku requests. Prompts longer than 100,000 tokens cost five times as much.
Those per-token reductions may not translate directly into the same savings for real workloads. Haiku 5.5 uses an updated tokenizer that consumes slightly more tokens per task than its predecessor. Anthropic has described a similar effect with the Opus 4.x models, where the tokenizer change alone increased token use by about 30 percent.
For buyers, the practical cost depends on both the price per token and how many tokens a task uses. The lower rates are clearest for the prompt lengths Anthropic says are common, but organizations will need to consider the updated token counts when estimating overall spend.
Benchmark gains are strongest in computer use
The reported scores show a substantial rise over Haiku 4.5 on several evaluations. On GDPval-AA v2.1, a knowledge benchmark, Haiku 5.5 scores 1,620, compared with 735 for its predecessor. On Humanity's Last Exam, it scores 45.9 percent without tools and 57.4 percent with tools, up from 10.2 and 18.7 percent.
Computer use stands out as a particularly large improvement. On OSWorld-2.1, where the model operates a computer, Haiku 5.5 scores 72.4 percent, compared with 15.7 percent for Haiku 4.5. The article notes that this kind of use can consume many tokens, making a lower-cost, fast model potentially useful for such workloads.
On Terminal-Bench 4.0, an agentic coding benchmark, Haiku 5.5 reaches 39.2 percent; Haiku 4.5 scored zero. Anthropic's comparison also puts Haiku 5.5 ahead of OpenAI's budget model GPT-6 Luna across every tested category. The reference scores for Sonnet 5.5 show that Haiku 5.5 still trails Anthropic's larger model.
Reasoning controls help match effort to the task
Haiku 5.5 is the first Haiku-class model with adjustable reasoning levels. That gives users a way to balance cost against quality, depending on what a task needs. Anthropic says the model is best suited to narrowly scoped work, including compaction, summarization, and sub-agent tasks.
The distinction matters for teams choosing among models. Haiku 5.5 is positioned for repeated, well-defined work where speed and cost are central concerns. For complex agentic coding, Anthropic continues to recommend Sonnet 5.5 and Opus 5.5.
The model's cybersecurity safeguards are tighter than those of its predecessor, while allowing a broader range of defensive tasks than Sonnet 5.5. Penetration testing remains blocked. Organizations with broader needs can apply to Anthropic's verification programs for life sciences and cybersecurity.
Availability and other pricing changes
Haiku 5.5 is available across platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic is also updating its Python and TypeScript SDKs with beta support for computer use and browser use.
The launch includes a separate price reduction for Sonnet 5.5: cache read costs are cut by 50 percent, from $0.20 to $0.10 per million tokens. Anthropic says the change should lower costs for most agentic tasks by about 20 percent. Taken together, the pricing updates give customers more room to use models for recurring work, while the benchmark and task guidance help clarify where the smaller Haiku model fits.
Anthropic is also introducing monthly API credits for some subscribers. Max-5x subscribers receive $100, Max-20x subscribers receive $200, and Team subscribers receive up to $500 per month. The credits can be used to experiment with tools, apps, and agents through the API.