What happened
On 7 October Anthropic released Claude Haiku 5.5, the new smallest model in its family. For short requests the price fell sharply: ten cents per million input tokens and fifty cents for output. Haiku 4.5 costs a dollar and five dollars. The context window grew to a million tokens, the longest response to 128 thousand, and Haiku gets effort controls for the first time, meaning you can trade reasoning quality for cost on each task.
The news is in the price list. Haiku 5.5 is billed by prompt length. Above 100,000 tokens, input costs fifty cents, output USD 2.50, and a cache read five cents instead of one. Every line is five times more expensive. The documentation adds two details that are easy to miss. Prompt length counts every input token, including those read from and written to the cache. And each request is priced on its own: if it crosses the threshold, it pays the higher rate for everything, response included. Anthropic's other models from the 4.6 generation onward bill the full million tokens at a single rate. Haiku 5.5 is the exception.
Anthropic says prompts of up to 100,000 tokens made up around nine in ten requests to the previous Haiku. It also says the new tokenizer uses slightly more tokens for the same task than Haiku 4.5, so part of the cut is eaten by the change in how tokens are counted.
Our take
The headline price is real, but it applies to requests, not to tasks. A single call, such as classifying an email, pulling fields from an invoice or summarising a document, sits well below the threshold and really does cost a fraction of the old price. An agent is different. When the price war moved into the cache, we wrote that an agent sends the model the whole conversation so far at every step. That conversation keeps growing, and at some step it passes 100,000 tokens. From then on, every following step pays five times more for everything, including cached context whose cost was supposed to be negligible.
A simplified calculation to make it visible. This is an illustrative example, not a measurement of any deployment. An agent session on Haiku 5.5 runs 40 steps. It starts with 20,000 tokens of context; each step adds 4,000 tokens and produces a 1,000-token response. Everything except the new fragment is read from the cache.
- If the whole session paid the below-threshold rate: about USD 0.08.
- Under the actual price list, where the last 19 steps cross the threshold: about USD 0.26, more than three times as much.
- The same session on Haiku 4.5: about USD 0.79.
Two things follow. Moving off Haiku 4.5 pays either way, because even above the threshold Haiku 5.5 costs half of what its predecessor does. But an agent budget built on the ten-cent rate is understated several times over. In this example, two thirds of the bill comes from context length alone, not from the choice of model.
We flag this as our interpretation: the threshold shows how the vendor sees the small model's role. Short, repeatable, high-volume tasks and subagent work, not long autonomous sessions. Anthropic itself recommends larger models for complex agentic coding. We expect prompt-length pricing to show up in more low-cost models, because it lets a vendor lower the headline rate without giving up margin on long context. That is a forecast, not a fact.
Why it matters
Private Equity
A portfolio company whose product runs on a cheap model at high volume can improve its margin on this release for real, or only in the deck. At the next review, ask for the distribution of request lengths: what share of traffic crosses 100,000 tokens, and what in the application keeps it from doing so. A company that only knows the average doesn't know which side of the threshold its cost sits on. It is one of the questions we start with in an AI cost review on the fund side: the distribution first, the price list second.
Enterprise
The threshold moves the cost decision from procurement into architecture. Three places are worth checking. First, agents with long histories: is context trimmed or summarised before it reaches the threshold. Second, document retrieval that pastes whole files into the prompt just in case, instead of the fragments it needs. Third, batch jobs. Batch halves the rates, but the threshold applies there too, so overnight processing of long documents may cost something other than the forecast assumes after the migration.
SMB / mid-market
Typical automations in a smaller company, such as inbox handling, ticket classification and pulling data from documents, fit under the threshold. For them this is simply good news. If someone built you an automation on Haiku 4.5, ask when they will switch it to 5.5 and whether they checked the results after the tokenizer change and at the chosen effort level. The trouble starts only when someone drops a whole binder of contracts into a single request.
One move for this week
Pull the distribution of prompt lengths for the last thirty days, from your provider's console or your application logs, for everything running on Haiku. Count what share of requests crosses 100,000 tokens and which processes send them. If they are single calls, switch without worry. If they are agents with a growing history, set up context trimming below the threshold first, and only then change the model. Otherwise the new price will only work in the spreadsheet. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI