Claude Haiku 5.5 cut small-model API prices by 90%. The catch is in the long prompts.

What happened

Anthropic released Claude Haiku 5.5 on October 7. The pricing is the news.

For prompts up to 100,000 tokens, the model lists at $0.10 per million input tokens and $0.50 per million output tokens, according to coverage of the launch. Haiku 4.5 listed at $1 and $5. That's a 90% cut on both sides.

The model is available through Anthropic's own platform plus Amazon Web Services, Google Cloud and Microsoft Azure, per the same reporting. Anthropic also said some subscribers get monthly API credits: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team plans.

One caveat up front. These figures come from secondary coverage, not from an Anthropic pricing page we could pull directly. Check the current docs before you build a budget on them.

What it actually costs

Run the math on a realistic month.

Say you push 10 million input tokens and get 2 million output tokens back. That's roughly 150 hours of podcast transcript cleaned up, chaptered and summarized, with headline variants on top.

Input: $1. Output: $1. Total model cost: $2.

That is the base model bill and nothing else. No hosting. No retrieval layer. No storage. No orchestration. No human editor reading the output. The $2 is real, and it is also the smallest line on the invoice.

The part that will surprise you

The headline price has a ceiling. Above 100,000 tokens in a single prompt, Haiku 5.5 costs $0.50 per million input and $2.50 per million output — five times the advertised rate.

That threshold is lower than it sounds. A two-hour interview transcript can run past 100,000 tokens on its own. Dump a full episode plus show notes plus your style guide into one call and you've crossed it without noticing.

Anthropic also lists separate cache-read and cache-write charges. If you're running the same archive query repeatedly — a common pattern for anyone building search over their back catalog — those are a real line item, not a rounding error.

Chunk your inputs. Keep each call under the threshold. That single habit is the difference between the price you read about and the price you pay.

What this is NOT

It is not a quality upgrade for your hardest work. Haiku is the small model in Anthropic's lineup. Cheap first-pass labor, not the thing you hand a script to.

It is not a reason to cancel another subscription. Nothing here says Haiku 5.5 replaces whatever you use for drafting or reasoning. The argument is that routine, high-volume tasks no longer need a premium model.

It is not confirmed production savings. No customer case study in the available reporting shows anyone actually cutting a bill by 90% in practice. The list price dropped. Your invoice depends on your usage shape.

And it is not a consumer price change. If you pay a flat monthly fee for a chat app, this does nothing for you. This is API pricing — it matters if you or a contractor are building something.

What to do about it

1. Inventory the AI tasks you currently run on a premium model that don't need one. Metadata, tags, transcript cleanup, thumbnail title variants, newsletter segmentation, archive classification. Those are Haiku jobs.
2. Measure your typical prompt size in tokens before you migrate anything. If you're routinely over 100,000, the 90% headline doesn't apply to you.
3. Split long transcripts into chunks under the threshold and stitch results. Test whether quality holds.
4. Check whether your Max or Team plan includes the API credits named above. If it does, your first month of testing may cost nothing.
5. Price cache reads and writes separately if you're building repeat queries over an archive. Model it; don't assume.
6. Verify every number here against Anthropic's live pricing documentation before you commit. The launch coverage is a few days old and pricing pages move.
7. Re-run the build-versus-buy question on any AI vendor charging you a per-seat fee for work that's now a $2 API call. Ask them what they're adding.