Claude Haiku 5.5: what to know about its features, prices and performance claims

Those are Anthropic’s suggested uses, not independent findings that Haiku will perform well on every task in those categories. The company says Sonnet 5.5 remains a better fit for more complex agentic coding, in which a model handles a sequence of steps toward a goal.

The release adds a developer-controlled effort level to the Haiku line. Anthropic’s platform documentation says adaptive thinking is enabled by default and the standard effort setting is medium. The listed context window is 1 million tokens—the units a model processes—and Haiku can produce up to 128,000 output tokens.

How much cheaper is Haiku 5.5?

For prompts up to 100,000 tokens, Anthropic lists Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens. Above that prompt-length threshold, the listed rates are $0.50 per million input tokens and $2.50 per million output tokens.

Anthropic says the rates for shorter prompts are 90% below Haiku 4.5’s prices of $1 per million input tokens and $5 per million output tokens. For longer prompts, it says Haiku 5.5’s rates are 50% lower. The company also says about 90% of Haiku 4.5 requests fell into the lower, up-to-100,000-token category.

Those rate cuts do not translate directly into the same percentage reduction in the cost of a completed task. A tokenizer breaks text into tokens for a model to process. Anthropic’s documentation says Haiku 5.5’s tokenizer produces about 30% more tokens from the same text than Haiku 4.5’s, so a customer’s workload may use more billable tokens even when the text itself is unchanged.

Taking the tokenizer difference into account, Anthropic estimates Haiku 5.5’s average running cost is about 75% lower than Haiku 4.5’s. That is the company’s estimate, not a guaranteed saving on an individual bill. The final cost also depends on prompt length, input and output volume, and whether a prompt crosses the 100,000-token pricing threshold.

A small hands-on comparison illustrates why token counts matter, but does not provide a universal conversion. Developer and writer Simon Willison reported that one long test prompt used about 1.25 times as many tokens with Haiku 5.5 as with Haiku 4.5. That was his measurement for that prompt; it is distinct from Anthropic’s approximate 30% tokenizer estimate for the same text.

What does Sonnet 5.5’s cache price cut change?

Anthropic lowered Sonnet 5.5’s cache-read price from $0.20 to $0.10 per million tokens, effective Wednesday. Prompt caching lets a system reuse input that has already been supplied; a cache read is the charge for using that stored input again. The read rate has been halved, but that does not mean the total cost of Sonnet 5.5 has been cut in half. Anthropic’s Sonnet update describes the pricing change.

Anthropic estimates the lower rate will reduce costs by around 20% on most agentic work. That estimate depends on the share of a workload made up of cached input. Tasks that use cache reads to different degrees can see different effects, so the company’s estimate is not a promised discount on every task or customer bill.

The Haiku and Sonnet changes are therefore different comparisons. Haiku’s listed rates cover input and output tokens, with a higher price tier for longer prompts. Sonnet’s reduction applies to cached input only. Customers comparing the models need to test them on the same work and account for token use and cache behavior, rather than comparing the headline rate reductions alone.

What do independent tests show about Haiku 5.5?

Anthropic’s launch page reports Haiku 5.5 ahead of GPT-6 Luna on two shared benchmarks: 72.4% to 48.9% on the offline subset of OSWorld 2.1, and 39.2% to 16.4% on Terminal-Bench 4.0. These are results published by Anthropic. The release-day reporting does not establish an independent, matched replication of those comparisons, so they do not settle how the models compare on customer workloads.

There are outside observations, but they are narrower than a matched benchmark. SiliconANGLE reported that Asana tested Haiku 5.5 before release and found lower latency than its current model in an internal agent evaluation. Aaron Vinh, a staff software engineer at Asana, described the experience to SiliconANGLE as “It’s a noticeably snappier experience.” That is a customer’s account of its own evaluation, not a standardized comparison.

Willison’s hands-on report describes an image-generation task involving a pelican and a bicycle, compared with an earlier Haiku 4.5 attempt. His token count for one long prompt is useful as an example of how usage can vary, but it does not establish a general performance ranking or a typical customer’s bill.

Haiku 5.5 is released and available, and the listed prices are specified. Whether it reduces the cost of a particular job—and how much the Sonnet cache cut saves—still depends on workload-specific measurements. Independent tests using matched prompts, tools, effort settings and scoring methods would be needed to verify broader performance comparisons.