Claude Haiku 5.5: Anthropic Cuts Prices by Up to 90%

Claude Haiku 5.5: Anthropic Cuts Prices by Up to 90%

Anthropic has launched Claude Haiku 5.5, which it calls its fastest and cheapest small model so far. The release also brings lower prices for Sonnet 5.5 and a new monthly API credit scheme for paying subscribers. Taken together, the announcements show that the price competition between AI labs is still intensifying.

Cheaper per token, with a catch

Haiku 5.5 is built for high-volume work where cost matters most. Anthropic lists summarization, database queries, classification and live customer support as typical uses.

The pricing is the headline. On average, Haiku 5.5 costs about 75 percent less than Haiku 4.5. For prompts of up to 100,000 tokens, prices fall by as much as 90 percent. Anthropic says that range covered roughly 90 percent of all earlier Haiku requests. Prompts above 100,000 tokens cost five times as much.

There is one caveat. Haiku 5.5 uses an updated tokenizer that needs slightly more tokens for the same task. Anthropic saw a similar effect with its Opus 4.x models, where the tokenizer change alone raised token usage by about 30 percent. Real savings will probably be smaller than the per-token figures suggest.

Big jumps on benchmarks

Compared with Haiku 4.5, the gains are large:

  1. GDPval-AA v2.1 (knowledge work): 1,620, up from 735.
  2. Humanity's Last Exam: 45.9 percent without tools and 57.4 percent with tools, up from 10.2 and 18.7 percent.
  3. OSWorld-2.1 (computer use): 72.4 percent, up from 15.7 percent.
  4. Terminal-Bench 4.0 (agentic coding): 39.2 percent, where Haiku 4.5 scored zero.

Computer use is the most notable result. In this mode the model operates a computer on its own, which consumes a lot of tokens. A fast, cheap model is a natural match for that kind of workload.

Anthropic also compares Haiku 5.5 with OpenAI's budget model GPT-6 Luna, and Haiku leads in every category tested. The Sonnet 5.5 reference scores show a clear gap, though. The small model is still well behind Anthropic's larger one.

Adjustable reasoning and tighter cyber rules

Haiku 5.5 is the first Haiku-class model with adjustable reasoning levels, so users can trade cost against quality. Anthropic recommends it for narrowly scoped jobs such as compaction, summarization or work as a sub-agent. For complex agentic coding, the company still points to Sonnet 5.5 and Opus 5.5.

Cybersecurity safeguards are stricter than on Haiku 4.5. At the same time, they permit a wider range of defensive tasks than Sonnet 5.5 does, partly because Haiku is less capable overall. Penetration testing remains blocked. Organizations with broader needs can apply to Anthropic's verification programs for life sciences and cybersecurity.

The model is available now on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.

Sonnet price cut and API credits

Anthropic is also halving the cache read price for Sonnet 5.5, from $0.20 to $0.10 per million tokens. The company expects this to reduce costs for most agentic tasks by about 20 percent. The timing points to a response to OpenAI's new GPT-6.1 series, which includes GPT-6.1 Sol, pitched at a fraction of its flagship's cost.

Subscribers will now get monthly API credits. Max-5x users receive $100, Max-20x users receive $200, and Team subscribers get up to $500 per month. The credits can be used to try out tools, apps and agents through the API. This follows Anthropic's earlier push with free credits for startups.

Anthropic is also updating its Python and TypeScript SDKs, adding beta support for computer use and browser use.

Our Take

For teams running agents at scale, this release matters more than another frontier model would. Much agent work is repetitive: summarizing context, classifying inputs, clicking through interfaces. A cheap model that handles these steps well can change the economics of a whole pipeline, especially when it serves as a sub-agent under a larger model.

The tokenizer caveat is a reminder to measure, not assume. Per-token prices are only half of the bill. Teams should compare full task costs before switching.

The bigger pattern is clear. Labs are competing hard on price, and Anthropic's Sonnet cut looks like a direct answer to OpenAI. That fits with signs that business AI spending is falling even as usage grows. It is worth watching whether OpenAI responds with its own cuts, and whether independent tests confirm Haiku 5.5's computer-use scores outside Anthropic's own benchmarks.