Claude Haiku 5.5 API: slower than Haiku 4.5, cheaper per answer
Claude Haiku 5.5 costs less per answer than Haiku 4.5 and takes about twice as long to reply. That is what we measured on 40 shared prompts the day it went live on FastMetal, under the model id anthropic-claude-haiku-5-5.
Anthropic announced Haiku 5.5 on October 7, 2026 for high-volume, cost-sensitive work: summarization, context compaction, database queries and classification. Anthropic says it runs cheaper than Haiku 4.5.
What is new
- Effort control. Haiku 5.5 is the first Haiku that lets the caller set how deeply it reasons, so you choose cost or quality per request.
- A 1M-token context, with image input.
- Tool calling and structured output. We called it through FastMetal and confirmed that a forced tool call (
tool_choicenaming the function) and a strict JSON schema both work. - OpenAI-compatible: change the
modelvalue in your existing code and it runs. - Your prompts are not used for training. The host's terms state that inference data is not trained on.
One difference from the model it replaces: anthropic-claude-haiku-4-5 on FastMetal runs in Japan (Tokyo), and Haiku 5.5 runs on a global route. If your data has to be processed in Japan, keep using Haiku 4.5.
How it is billed
Input, output, cache reads and cache writes are each billed per million tokens, from your prepaid balance.
Haiku 5.5 is the one current Claude model whose rates depend on prompt length. When a single request's prompt (cache reads and writes included) is over 100,000 tokens, every line of that request is billed at the higher tier. For long documents, splitting the input into requests under 100,000 tokens can cost less.
Current rates are on the Claude Haiku 5.5 model page and the pricing page.
Haiku 5.5 vs Haiku 4.5, measured
We sent the same Japanese and English prompts to both models with FastMetal's comparison runner on October 8, 2026. The table covers the 40 prompts both models answered, as medians per answer.
| Model | Latency | Output tokens | Answer length (characters) |
|---|---|---|---|
| Claude Haiku 5.5 | 11.2 s | 2,115 | 2,192 |
| Claude Haiku 4.5 | 5.6 s | 788 | 1,998 |
What the numbers show:
- Haiku 5.5 is slower. It was faster on only 7 of the 40 prompts. It reasons before it answers, and that takes time.
- It produces about 2.7 times the output tokens for answers of about the same length. The extra tokens are hidden reasoning. The visible answer was slightly longer from Haiku 5.5 on 27 of the 40 prompts.
- It still costs less per answer. Its output rate is far below Haiku 4.5's, so on output alone each answer came out cheaper, which matches Anthropic's claim. Compare the rates on the Haiku 5.5 and Haiku 4.5 pages.
For real-time work where speed comes first, Haiku 4.5 can still be the better fit. For large batch jobs, Haiku 5.5's lower cost per answer adds up. Our earlier Claude Haiku 4.5 review covers the older model in more depth.
Calling it
Point any OpenAI-compatible client at FastMetal and set the model:
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic-claude-haiku-5-5",
"max_tokens": 16000,
"messages": [
{"role": "user", "content": "Classify this support ticket as billing, bug or other: I cannot log in."}
]
}'
Always set max_tokens. Without it the limit is 4,096 tokens, and because Haiku 5.5 spends tokens on reasoning, a long answer can hit the limit before any text comes out. The tokens it used are still billed. In our comparison run, one prompt that asked for a long how-to guide came back empty for exactly this reason.
Where it fits
Pick Haiku 5.5 when cost per request matters more than response time. If you run classification or summarization at volume, send the same prompts to Haiku 5.5 and Haiku 4.5 and compare. Everything we serve is in the model catalog.