Claude Haiku 5.5 API pricing and cost calculator

How this is calculated

Start with a summarization job that sends 20,000 input tokens per call and gets 1,000 back. With half the input read from cache, one call costs $0.0016: $0.0011 for input and $0.0005 for output. With no cache hits the same call costs $0.0025, and with the whole prompt read from cache it costs $0.0007. Over 10,000 calls a month, the chart's starting total is $16, shown to the cent as $16. A subagent fleet looks different. Say each helper sends 8,000 input tokens, returns 600, and reads most of its prompt from cache because the instructions and tool list never change. A fleet that makes 200,000 of those calls costs $90.40. Browser use is the heavy case. A page state, a screenshot and the history so far can add up to 40,000 input tokens per step, and a task might take 50 steps with 800 output tokens each. Across 1,000 such tasks with three quarters of the input cached, that's $85. How many tokens a screenshot takes depends on your setup, so count a few real sessions before you trust that figure. Output bills at $0.50 per million and uncached input at $0.10, so check how much of each call's cost comes from the reply before you trim the prompt. Prompt length changes the rate. Anthropic moves the whole request to a higher tier when a prompt goes over 100,000 tokens, and prompt length includes cache reads and cache writes. Above that line the rates are $0.50 input, $0.05 cache reads and $2.50 output per million. A call with 150,000 input tokens and 2,000 output tokens, half of it cached, costs $0.04625 at whichever tier that length lands in. The context window is 1,000,000 tokens and the maximum output is 128,000 tokens. Input can be text, images or files.

Verdict

Haiku 5.5 is built for work you run thousands of times, so measure the cost of one finished job and not the price of one call. Log input tokens, cache read tokens, output tokens and retries for a sample of your real traffic, then price the sample with the calculator. Keep an eye on prompt length too, because one request that crosses 100,000 tokens is billed at the higher tier from its first token. Cache writes, tool fees and failed attempts sit outside the per call examples, so add them from your own logs before you commit to a budget. If your quality bar needs a bigger model on some of the work, route only those calls there and let Haiku take the rest.