Claude Haiku 5.5: What Switching Really Saves and Which Settings Now Throw Errors
Claude Haiku 5.5 has been Anthropic's smallest model since October 7, 2026 and costs 90% less per token than Haiku 4.5. Because the same text produces about 30% more tokens and requests over 100,000 tokens are billed entirely at a higher rate, you really save closer to 85% on short tasks and only about a third on long documents. Workflows that set a temperature, prefill the response, or always read the first content block will break.
What is Claude Haiku 5.5 and what does it cost?
Since October 7, 2026, Claude Haiku 5.5 has been Anthropic's smallest and fastest model, built for large volumes of short tasks: sorting requests, pulling fields out of documents, routing tickets (model overview). You pay per token, the small chunks of text a model splits language into. According to the pricing page, there are two tiers depending on how long your request is:
| Price per million tokens | Haiku 4.5 | Haiku 5.5 up to 100,000 tokens | Haiku 5.5 over 100,000 tokens | Sonnet 5.5 |
|---|---|---|---|---|
| Input | $1.00 | $0.10 | $0.50 | $2.00 |
| Output | $5.00 | $0.50 | $2.50 | $10.00 |
| Cache read | $0.10 | $0.01 | $0.05 | $0.10 |
Through the Batch API, where requests are collected and processed later, all prices are cut in half. The model handles text and images, works with up to 1 million tokens of context, and has a knowledge cutoff of June 2026. It runs on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry (model page). According to Make, you can also use it in Make since October 10 (Make Community).
Performance takes a big jump. On operating software (OSWorld 2.1), Haiku 5.5 scores 72.4% according to Anthropic, versus 15.7% for Haiku 4.5. For work in the terminal, though, Sonnet 5.5 at 70.6% is far ahead of Haiku 5.5 at 39.2%, and Anthropic itself still recommends Sonnet or Opus for complex coding tasks.
How much cheaper is it really?
About 85% on short tasks, but only about a third on long documents. The 90% applies per token, and three things shift the math:
- More tokens for the same text. Haiku 5.5 uses a new tokenizer, so the same text produces about 30% more tokens than on Haiku 4.5 (migration guide).
- The 100,000 token threshold applies to the whole request. If the input goes over it, everything is billed at the higher rate, including cached parts (pricing page). A contract that was 80,000 tokens on Haiku 4.5 lands at around 104,000 tokens and crosses the line.
- Images get sharper and therefore pricier. A photo of 2,000 by 1,500 pixels takes about 2.5 times as many image tokens as on Haiku 4.5.
Our math, using a factor of 1.3 for the tokenizer:
| Case | Savings vs. Haiku 4.5 |
|---|---|
| Text, request under 100,000 tokens | about 87% |
| Large image (2,000 x 1,500 px) | about 75% |
| Text, request over 100,000 tokens | about 35% |
Then there is thinking. By default, Haiku 5.5 reasons before it answers, and those thinking tokens count toward the output limit max_tokens (effort docs). An example with assumed values: you classify 10,000 emails a month, each with 2,000 input tokens and a 150 token answer, as measured on Haiku 4.5. On Haiku 4.5 that costs about $27.50. On Haiku 5.5 it's $3.58 without thinking and about $5.10 if each email uses 300 thinking tokens. Still more than 80% less, but not 90%. Whether an open-weight model costs even less for the same tasks is something we worked out in our DeepSeek V4.1 Flash vs. Haiku 5.5 comparison.
Which settings break when you switch?
Five things that are often set in automations lead to errors or empty answers, according to the migration guide:
- Temperature: Leave out
temperature,top_p, andtop_k. The only allowed value istemperature1, and anything else returns a 400 error. If you used temperature 0 for stable classification, you now get an error instead. - Prefilled response (prefill): If your request ends with the start of an assistant message, such as an opening bracket for JSON, Haiku 5.5 rejects it. For fixed formats, use structured outputs or a tool with a fixed set of allowed values instead. On Amazon Bedrock, structured outputs are not available for Haiku 5.5.
- First content block: Because a response can start with a thinking block, a fixed mapping to the first content block, for example in Make's HTTP module, suddenly returns no text. Select blocks by their type.
- Tight
max_tokens: If the limit is small, the model can use it up while thinking and stop without any answer text. Raise the limit or lower the effort. - Refusals: Haiku 5.5 has safety classifiers that can end a request with
stop_reasonrefusal. Your workflow needs an error branch for that.
Which effort level should you set?
Use low for simple bulk tasks and start with the default medium for everything else. The effort setting controls how much the model thinks, and with it quality, speed, and cost. According to Anthropic, low fits short tool tasks and simple high-volume requests, while high fits knowledge work and strict instruction following. Up to high, you can also switch thinking off entirely.
What does this mean for you?
If you use Claude in Make or n8n to sort, extract, or route, switching almost always pays off. Our take, and how we handle it in client projects:
- Find what you run today: Which scenarios use Haiku 4.5 or still Sonnet 4.5? Anthropic retires Sonnet 4.5 on November 30, 2026, and the replacement is Sonnet 5.5 (model deprecations). Haiku 4.5 is still active, and its earliest possible retirement date is October 15, 2026. There is no deprecation notice yet, and Anthropic announces retirements at least 60 days in advance. So Haiku 4.5 will keep running on the Claude API until at least mid-December 2026.
- Test a copy: Clone the scenario, swap the model, remove temperature and prefill, read the response by block type, and check 20 to 50 real cases against the old results.
- Set the effort: Start with
lowfor simple bulk tasks andmediumotherwise, and only raise it where results get worse. - Stay under 100,000 tokens: Split long PDFs first, or your savings shrink to about a third.
- Measure costs again: After a week, look at the actual token counts instead of projecting the old ones.
What Make and n8n cost to run is covered in our Make vs n8n comparison. Whether running your own model beats an API is something we worked through using Aleph Alpha Kolibri as the example. If you want your AI workflows reviewed or migrated, take a look at our services or get in touch.
Frequently asked questions
Do multi-turn chats get more expensive with Haiku 5.5?
Yes, they tend to. Thinking blocks from earlier responses stay in context and count as input tokens, while Haiku 4.5 kept only those from the latest turn. You can remove older thinking blocks with thinking block clearing.
Can I see what Haiku 5.5 is thinking?
Not by default: thinking blocks come back with an empty thinking field and only a signature. With the display setting summarized, you get a summary of the reasoning.
Does computer use work with Haiku 5.5 the way it did before?
No. The old computer_20250124 tool returns a 400 error on the Claude API and Google Cloud, where you need the new computer_toolset_20260801 toolset. On Amazon Bedrock, use computer_20251124 with its own beta header.
Does Haiku 5.5 cost more if my data is processed only in the US?
Yes. With US-only inference, every token type costs 1.1 times as much, including requests over 100,000 tokens.
Sources
Tell us what you're planning. You'll get an honest assessment within 24 hours.
More articles