Claude Haiku 5.5: price, speed and when to use it

What is new in Claude Haiku 5.5

Anthropic calls Haiku 5.5 its cheapest, fastest and most capable small model so far. It is the first Haiku-class model with an adjustable effort setting, from Low to Max, so developers can trade cost and latency against capability.

Pricing starts at US$0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens, rising to $0.50 and $2.50 for longer requests. Anthropic puts the saving versus Haiku 4.5 at about 75% on average: 90% for requests up to 100K tokens and 50% for longer ones.

Sources: Anthropic: Introducing Claude Haiku 5.5

Benchmarks: a big jump over Haiku 4.5, still below Sonnet

Anthropic reports 72.4% on the offline subset of OSWorld 2.1, against 15.7% for Haiku 4.5, and 45.9% on Humanity's Last Exam without tools, against 10.2%. On Terminal-Bench 4.0 it scores 39.2%, where Haiku 4.5 scored 0%.

In the same tables Sonnet 5.5 scores higher on every benchmark listed, and Anthropic says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks.

Sources: Anthropic: Introducing Claude Haiku 5.5

Recommended uses and the limits Anthropic lists

Anthropic recommends Haiku 5.5 for high-volume, cost-sensitive work such as summaries, database queries and classification; as a sub-agent working alongside Opus 5.5 or Sonnet 5.5; and for speed-sensitive tasks such as live customer support and browser use.

It also lists limits: Haiku 5.5 uses a new tokenizer, so a task consumes slightly more tokens; its cybersecurity safeguards are stricter than Haiku 4.5's; and it runs slower than the Opus models in Fast Mode.

• Good fit: summaries, tagging and routing, extraction, short support replies.
• Good fit: sub-tasks that a larger model plans and checks.
• Weaker fit: long, multi-step coding work that needs careful planning.

Sources: Anthropic: Introducing Claude Haiku 5.5

Haiku 5.5 or Sonnet: how to decide

The price gap makes Haiku attractive for anything you run thousands of times. As simple arithmetic from Anthropic's published rates: one million short requests of 2,000 input tokens each is 2 billion input tokens, about US$200 on Haiku 5.5 at the under-100K rate before output costs, against ten times that at Haiku 4.5's old input price.

Our rule of thumb is to start with the cheapest model that passes a test set of your own examples, and move up only where it fails. Keep the failures: they show exactly which tasks deserve Sonnet.

Test a small model on a real high-volume task

Collect ten real examples of the task you want to automate, such as customer messages, short documents or support tickets, with the answer you consider correct. Run the prompt below and compare the labels with your own. If most match and the misses are easy to catch, a small model is probably enough. Remove personal data before pasting anything.

Classify each message below into one of these categories: [category list]. For each message return: number, category, confidence (high, medium or low) and a one-line reason. If a message fits no category, say so instead of guessing. Messages: [paste 10 messages, one per line]