Claude Haiku 5.5 Costs 90% Less, Until Your Prompt Hits 100,000 Tokens. Then the Price Jumps Fivefold.

Anthropic's Claude Haiku 5.5 is now available through Amazon Bedrock at $0.10 per million input tokens and $0.50 per million output tokens, but only for requests under 100,000 input tokens. At or above that threshold, the listed rates are $0.50 and $2.50 per million tokens, a fivefold increase on both input and output.

That makes the headline price conditional. The low rate applies to short-context work. An application that sends large document collections, long conversation histories, or heavy retrieved context with every request could land in the higher tier on a regular basis. For those workloads, the number that matters is the average input size per request, and how often it crosses the 100,000-token line.

The savings claims come from the vendors themselves. Anthropic says Haiku 5.5 costs 90% less than Haiku 4.5 below the 100,000-token threshold, and calls it its fastest, cheapest, and most capable small model. AWS confirms Bedrock availability, describes it as Anthropic's most capable Haiku for coding, tool use, computer use, and agentic tasks, and puts the saving at about 75% for most tasks. The two figures aren't the same claim. One is a rate comparison under a specific threshold, the other a broader estimate across typical use, and neither has been independently tested.

What's missing is the evidence that would turn a price list into a cost estimate. The available material has no independent benchmark replication, no completed-task cost studies, and nothing showing the savings hold for workloads that regularly cross into the higher tier. Real spend also depends on output length, retries, and how many model calls a task takes. A cheaper model that needs more attempts can erase part of the discount.

Bottom line: Haiku 5.5 is a real price cut with a real condition attached. If your requests stay comfortably below 100,000 input tokens, the listed rates are very low. If they don't, the fivefold step-up is what you'll actually be paying. Before switching, measure your typical input size, then compare total cost per completed task rather than the headline rate.