Claude Haiku 5.5 cuts Anthropic's small-model price by 90% for prompts under 100,000 tokens

Haiku 5.5 is Anthropic's new small model, with the ID claude-haiku-5-5 and no date suffix. Anthropic calls it its fastest model at standard speed and says that on average it costs around 75% less to run than Haiku 4.5 once the new tokenizer is taken into account. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, and GitHub Copilot added it the same day for its Pro, Pro+, Max, Business, and Enterprise plans.

Two price tiers, split at 100,000 tokens

Pricing depends on prompt length. Up to 100,000 tokens, input costs $0.10 per million tokens and output $0.50; cache reads are $0.01 and cache writes $0.125. Over 100,000 tokens the same four prices are $0.50, $2.50, $0.05, and $0.625. Haiku 4.5 charges $1 per million input tokens and $5 per million output tokens at any length, so the cut is 90% for short prompts and 50% for long ones.

Anthropic says 90% of requests to Haiku 4.5 stayed under the 100,000-token line, and its headline figure of a 75% average saving already accounts for the new tokenizer. Haiku 5.5 uses the same tokenizer as Claude 4.7 and later models, so the same text counts as approximately 30% more tokens than on Haiku 4.5. Prompts need recounting, and a max_tokens limit tuned for the old model may cut output short, because thinking tokens now count toward it.

What breaks when you swap the model ID

The context window grows from 200k to 1M tokens and the maximum output from 64k to 128k. Adaptive thinking is on by default, with effort levels from low to max and a default of medium, and a response can begin with thinking blocks even when the request never mentions thinking. The migration guide lists the requests that now fail:

• Manual extended thinking with budget_tokens returns a 400 error; use thinking: {"type": "adaptive"} and the effort parameter instead.
• Any temperature other than 1, any top_p other than 0.99, and any top_k value return a 400 error.
• A final assistant turn used as a prefill is rejected, even with thinking turned off.
• Computer use on the Claude API and Google Cloud works only through the computer_toolset_20260801 toolset.
• Thinking blocks are bound to the account that produced them and to an unchanged conversation prefix.
• Priority Tier is not supported on Haiku 5.5, so a capacity commitment on Haiku 4.5 does not carry over.

What Anthropic's benchmarks claim

According to the company's table, Haiku 5.5 scores 72.4% on the offline subset of OSWorld 2.1, against 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna, and 83.9% for Sonnet 5.5. On Terminal-Bench 4.0 it reaches 39.2%, where Haiku 4.5 scored 0.0%, GPT-6 Luna 16.4%, and Sonnet 5.5 70.6%. On Humanity's Last Exam without tools it reports 45.9%.

Anthropic itself draws the line: Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding, and Haiku 5.5 is meant for narrow, high-volume work such as compaction, summarization, classification, and subagent tasks. No independent measurement has been published yet, so every figure here is the vendor's own.

Cheaper cache reads on Sonnet 5.5, and API credits for subscribers

Alongside the launch, cache reads on Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens. Because cached prefixes make up a large share of agent traffic, Anthropic estimates that most agentic work on Sonnet 5.5 now costs around 20% less.

Anthropic is also rolling out a monthly API credit this week: Max 5x subscribers get $100, Max 20x subscribers $200, and Team subscriptions up to $500 pooled across their users. The credit can be spent on any model. The Python and TypeScript SDKs add computer use and browser use in beta at the same time.