Anthropic Prices Claude Haiku 5.5 Up to 90% Below Haiku 4.5

Anthropic released Claude Haiku 5.5, its new small model for high-volume work, on 7 October 2026 at US$0.10 per million input tokens and US$0.50 per million output tokens for prompts of up to 100,000 tokens. Those rates are 90% below the US$1.00 and US$5.00 charged for Claude Haiku 4.5, according to Anthropic's launch page. Longer prompts cost US$0.50 and US$2.50, a 50% cut.

Small models often do the repetitive work in a company's AI systems: sorting customer messages, summarising documents, answering routine chat. Anthropic aims Haiku 5.5 at those high-volume jobs. For a Thai company that pays for AI by the token, a job once too expensive to run on every customer message, such as tagging each chat by topic, may now fit the budget.

What a token is, and why the 100,000-token line matters

A token is a small piece of text, often a short word or part of a longer one. The model reads a prompt as tokens and writes its answer as tokens, and the bill counts both, at different rates.

The 100,000 line refers to the prompt: everything sent in, attached documents included. A chat message sits far below it; a stack of long contracts can pass it. Around 90% of requests to the previous Haiku were under the line, Anthropic says on its launch page. Once a prompt goes over, the whole request is billed at the higher rates, according to the Amazon Bedrock model card. The model's context window, prompt and answer combined, is 1 million tokens, Anthropic's documentation says.

Why the average saving is about 75%, not 90%

Anthropic's own average is lower: around 75% less to run. Its footnote names two reasons.

First, the roughly one request in ten that passes 100,000 tokens gets only 50% off. The footnote does not spell out the arithmetic, but each of those prompts is by definition longer than any prompt under the line, so the group's share of the bill is bigger than its share of requests.

Second, Haiku 5.5 uses a new tokenizer, the tool that cuts text into tokens. The launch page says only that the model uses "slightly more tokens per task". The documentation gives a figure for the tokenizer alone: the same text counts as about 30% more tokens than on Haiku 4.5. Anthropic has not said how many more tokens a typical task uses overall, a gap PPC Land also noted.

An SV News calculation from the published prices shows the effect. Take a support chatbot handling 1 million messages a month, each with 2,000 tokens in and 500 out:

• Haiku 4.5: 2,000 million input tokens at US$1.00 per million plus 500 million output tokens at US$5.00 per million is US$2,000 + US$2,500 = US$4,500.
• Haiku 5.5, same tokens: US$200 + US$250 = US$450, or 90% less.
• Haiku 5.5, 30% more tokens: US$450 x 1.3 = US$585, or 87% less.

The example assumes no caching or batch discounts, no prompts over 100,000 tokens, and 30% more tokens on both input and output. Long prompts, which get only 50% off, would pull the saving down towards Anthropic's 75% average.

An effort setting and Anthropic's own test scores

Haiku 5.5 is the first Haiku with an adjustable effort setting, which lets developers choose whether to optimise for lower cost or for more intelligence.

On Anthropic's own tests, Haiku 5.5 scored 72.4% on the offline subset of OSWorld 2.1, a test of how well an AI agent operates a computer, against 15.7% for Haiku 4.5.

Rogo, one of six customers quoted on Anthropic's launch page, said a Haiku 5.5 subagent pulls segment revenue from a company's annual report while a bigger model builds the slides.

It's accurate enough that we'd trust it there, and fast and cheap enough that we can run it a lot.

Haiku 5.5 and AWS's Thailand Region

Amazon Web Services (AWS) has offered older Claude models (Opus 4.6, Sonnet 4.6 and Haiku 4.5) to customers in its Thailand Region, ap-southeast-7, through global cross-Region inference on Amazon Bedrock since 24 February 2026, AWS said. Under that system a request can be processed in any of more than 20 AWS Regions, while logs and stored configurations stay in the Thailand Region.

Nothing confirms that Haiku 5.5 is processed in Thailand. Anthropic's launch page does not mention the country, and the only reference SV News found is one row in the AWS model card for Haiku 5.5, checked on 22 October 2026. It lists ap-southeast-7 only as a starting point for global routing and marks in-region processing as not supported. A Thai company that sends requests from the Thailand Region may have them processed abroad.

Sonnet 4.5 retires on the Claude API on 30 November

Haiku 4.5 users face no deadline yet on the Claude API. Anthropic's model deprecations page lists the model as active, retiring "not sooner than" 15 October 2026, a date that has passed with no retirement set.

The nearest fixed date is for an older model. Anthropic deprecated claude-sonnet-4-5-20250929 on 30 September 2026 and named claude-sonnet-5-5 as its replacement. Amazon Bedrock and Google Cloud set their own schedules, but on the Claude API, requests to Sonnet 4.5 will fail after 30 November 2026.

FAQ

How much does Claude Haiku 5.5 cost?

US$0.10 per million input tokens and US$0.50 per million output tokens on prompts of up to 100,000 tokens. If a prompt is longer, the whole request is billed at US$0.50 and US$2.50.

Why does Anthropic say 75% and not 90%?

Prompts over 100,000 tokens get only a 50% cut, and Haiku 5.5's new tokenizer uses more tokens for the same work. Anthropic's documentation puts the tokenizer effect at about 30% more tokens for the same text.

Can companies in Thailand use Haiku 5.5 through AWS?

Not as an in-country service. Nothing confirms Haiku 5.5 is processed in Thailand: AWS's model card lists its Thailand Region only as a starting point for global routing, so requests may be processed in another country.

Do Haiku 4.5 users get the lower prices?

No. The new rates apply only to Haiku 5.5, so a company has to switch models to pay them. Haiku 4.5 is still active on the Claude API with no retirement date set, and Anthropic promises at least 60 days' notice before retiring a public model.

Sources

• Anthropic: Introducing Claude Haiku 5.5
• Claude Docs: Claude Haiku 5.5 overview
• Claude Docs: Model deprecations
• AWS: Claude Haiku 5.5 model card, Amazon Bedrock User Guide
• AWS Machine Learning Blog: Global cross-Region inference for latest Anthropic Claude Opus, Sonnet and Haiku models on Amazon Bedrock in Thailand, Malaysia, Singapore, Indonesia, and Taiwan
• PPC Land: Anthropic Haiku 5.5 costs 90% less than 4.5 for prompts to 100,000 tokens