Claude Haiku 5.5 got cheaper per word.

Anthropic put out a new budget AI model last week and says it's 75% cheaper to run. Someone outside the company did the math on 3 real tasks. On 2 of them, the new model cost more per job than the old one. On 1, more than 3x as much.

Cheaper. Sure.

What launched

On October 7, Anthropic released Claude Haiku 5.5. It's the small, fast, cheap one. You probably don't chat with it. It lives inside apps and automations, doing thousands of tiny jobs nobody looks at. Anthropic's page says that "on average, it now costs around 75% less to run" than the old version, Haiku 4.5.

The price list looks even better than that. AI companies charge by the token. A token is a chunk of text, usually part of a word.

The old model charged $1 per million tokens you sent in and $5 per million it sent back. The new one charges $0.10 and $0.50. That's a 90% price cut. Big requests (over 100,000 tokens, roughly a 200-page book) get a smaller cut, 50%.

How a 90% price cut makes some jobs cost more

Someone priced out 3 standard tests that researchers use to compare AI models, old model versus new, cost per attempt.

On OSWorld, at the highest effort setting, the cost dropped from $1.45 to $0.61. Good.

On GDPval-AA, it went from $0.24 to $0.87.

On Humanity's Last Exam, from $0.09 to $0.21.

By my math, those last 2 are about 3.6x and 2.3x more expensive than before.

Two things changed besides the price tag.

The new model counts text differently. The same sentence comes out to roughly 25 to 30% more tokens than before. So you pay a lower price, on a bigger pile of tokens. The pile also got measured with a different ruler, which is convenient for the ruler's owner.

The new model always "thinks" before it answers. Developers can tell it how hard to think, from low up to max. They can't tell it to skip thinking. Thinking is more tokens, and tokens are the bill.

Nobody has published how much of the jump comes from the counting change and how much from the thinking. I'm not going to guess. What I can say is that "90% cheaper per token" and "more expensive per job" are both true at the same time, for some work.

About that 75%…

Anthropic's footnote says 90% of old-model requests were small, and that its math accounts for the new model using a different number of tokens. Anthropic doesn't say how it weighted requests, how much text was going in versus coming out, or how big a typical request was. So nobody outside can check the 75% against the price table sitting right next to it. Trust us, it's cheaper.

Here's the part I'd remember. "90% of requests" counts requests. It doesn't count dollars. 1 huge request can cost more than 500 tiny ones. So the other 10% of requests could be a big chunk of what people actually pay, and those only got the 50% cut.

My read: the 75% is a marketing number. For short, simple requests, which is most of what this model does, the savings are real and big. A per-token price still can't tell you what your bill does, and neither can an average built from someone else's requests. The only cost number worth trusting is yours.

Who should care

If you pay a flat monthly fee for ChatGPT, Claude or Copilot: nothing about your costs changed last week. You can stop reading. Go enjoy your Monday.

If someone built you something that calls an AI model: keep going. That covers a Zapier step, a Power Automate flow, a chatbot on your website, or a custom tool a freelancer or agency built.

Made-up example, but it makes the point. You pay a freelancer to build a flow that reads each vendor invoice and pulls out the total and due date. They see "90% cheaper," flip the flow to the new model, and move on. Great work, everyone. Nobody checks what a run costs afterward. If your invoices behave like the 2 tests that got pricier, the bill goes up, and the first person to notice is whoever opens the card statement.

The 1 question to send this week

Send whoever built your automation this: "What does 1 run cost right now, and did it change when you moved to the new model?" Ask for the answer in dollars per run. Dollars. Per run.

If they can answer, good. Put that number somewhere you'll see it every month.

If they can't, you just found the bigger problem. An automation nobody can price is an automation nobody's watching.

If you've got AI running inside your business and can't say what any of it costs per job, tell me what you've got running. I'll look at the process with you and find the numbers that matter.

Sources: Anthropic, Claude Haiku 5.5 announcement, October 7, 2026. braindetox.kr, Claude Haiku 5.5 pricing analysis. Simon Willison, Claude Haiku 5.5 write-up, October 7, 2026.