Claude Haiku 5.5 is the first time the cheap Claude tier looks like a model you'd actually route traffic to. On prompts up to 100,000 tokens, Anthropic prices the Oct. 7 model 90% lower than Haiku 4.5.
Short prompts are cheap. Long ones aren't. Stay under 100,000 tokens and you pay the low rate. Send a long document, a heavy tool trace, or a chat that's already piled up, and you step up to the other rate.
The rate card, by prompt length
| Model | Input per million tokens | Output per million tokens | Cache reads per million tokens |
|---|---|---|---|
| Haiku 5.5, prompts up to 100k | $0.10 | $0.50 | $0.01 |
| Haiku 5.5, prompts over 100k | $0.50 | $2.50 | $0.05 |
| Haiku 4.5 | $1.00 | $5.00 | $0.10 |
| Sonnet 5.5 | $2.00 | $10.00 | $0.10 |
Haiku 4.5 charged one rate at every length, and these rows replace that card. Sonnet's input and output stay higher, and its cache reads are now $0.10 per million tokens, down from $0.20.
A bigger context window doesn't keep you on the cheap row. The model docs give Haiku 5.5 a 1 million token window, and the same step-up after 100,000 tokens. The same text also counts as about 30% more tokens than it did on Haiku 4.5, so a prompt that used to sit under the line can cross it. You can send a long prompt. You leave the first row when you do.
The length that picks the row is the whole request, not the question someone just typed. A standing system prompt, a tool trace, and earlier turns all count toward 100,000 tokens. Sort a week of calls by that full length and you get two piles: traffic on the low rate, and calls that are already on the second row. If the standing prompt fills the cheap row, the new message isn't what made the call expensive.
If the same instructions ride along on every call, read the cache-read column instead of the input column. That cell is a different price on every row. On Sonnet, it's the one that just changed, and most agentic work on that model now runs about 20% cheaper. The input price didn't move.
Batch requests are half off input and output if the work can wait. That discount doesn't pick the row. Prompt length still does.
Where the small model earns the call
Haiku 5.5 is the one you point at work that repeats and has to come back quickly. It can sit beside Sonnet or Opus and take the narrow piece of a coding task while the larger model holds the plan. Leave complex agentic coding, the multi-step terminal work Terminal-Bench measures, on Sonnet or Opus.
It's the first Haiku-class model with an effort setting, and thinking is on by default. Effort is how you steer that toward a cheaper answer or a more careful one without swapping models. Drop it when you want the bill to stay small. Raise it when the same call has to be more careful. You stay on Haiku either way. At standard speed it's the fastest Claude shipped so far. Opus in Fast Mode still outruns it, so it isn't the fastest Claude in every mode.
Aaron Vinh, a staff software engineer at Asana, said that against the model they use today, "we saw over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn." That was their eval of an agent that triages bugs, sets up projects, and searches portfolios.
You can call it now on the Claude Platform as claude-haiku-5-5, and on Amazon Bedrock, Google Cloud, and Microsoft Azure. Max and Team accounts get a monthly API credit this week, from $100 up to $500 depending on the plan, with the Team amount pooled, and you can spend it on any model.
Point the classifiers, routers, extractors, and all-day subagents at Haiku 5.5, and watch whether those prompts stay on the shorter row. Leave the agent that has to finish a long coding job on Sonnet or Opus. If you have Max or Team, spend the new monthly credit on one real route before you change the default model in production.
Photo: Marc Mueller / Wikimedia Commons, CC0.
.jpg){kind=link}