What happened
On October 7 Anthropic launched Claude Haiku 5.5, its small, fast model. The interesting part isn't that it beats the previous one, although it does. It's the price.
For requests under 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1 and $5. That's 90% less. If a request goes over 100,000 tokens, the price rises to $0.50 and $2.50, still half of what it was. Anthropic estimates that, on average, a real workload comes out 75% cheaper.
The round number is no coincidence. On September 22 OpenAI had cut GPT-6 Luna, its budget model, to exactly $0.10 and $0.50. And this same week GPT-6 reached ChatGPT on every plan, including the free one, with the Luna version.
The numbers worth knowing
A token is a piece of a word. To get a feel for it, a normal email runs to a few hundred tokens and a twenty-page contract can go past ten thousand.
- Claude Haiku 5.5: $0.10 per million input tokens and $0.50 output below 100,000 tokens per request. Above that, $0.50 and $2.50.
- Claude Haiku 4.5, the previous model: $1 input and $5 output, whatever the size.
- GPT-6 Luna: $0.10 and $0.50 since September 22, when OpenAI halved its price. The size surcharge starts at 272,000 tokens.
- Cache: reading cached content costs $0.01 per million on Haiku 5.5. Handy when you always send the same long instructions.
- According to Anthropic, about 9 in 10 requests made to Haiku 4.5 fall into the cheap bracket.
Why it matters
Two weeks ago, in our September news roundup, we said the price floor was around $0.10 per million tokens. Now that floor is set by the two providers businesses use most, with models that work well on real tasks and not just in lab tests.
Anthropic says Haiku 5.5 is built for repetitive work: summaries, classification, database queries, live customer support and tasks that a large agent hands off to smaller ones. That is exactly what a small business usually wants to automate first. It is also the first Haiku with an effort setting, so you can decide task by task whether you want it to think harder or cost less.
Our take: price per task is no longer a reason to do nothing.
What it means for a small business: three quick sums
We ran the numbers with Haiku 5.5's published prices for three tasks we see often. They are our own estimates, with reasonable text sizes, in dollars. What matters is the order of magnitude.
- Classifying 3,000 emails a month and routing them, at about 1,500 input tokens and 200 output tokens per email: around $0.75 a month. With Haiku 4.5 it would have been around $7.50.
- Extracting data from 1,000 text invoices, at about 3,000 input tokens and 300 output tokens each: under half a dollar a month.
- Summarising 500 sales calls and pushing them to the CRM, at about 8,000 input tokens and 500 output tokens per call: around $0.53 a month.
Where the real cost sits now
If tokens cost cents, why isn't an AI project free? Because the cost has moved.
What costs money is designing the task well, connecting the model to your email, your ERP or your CRM, reviewing the cases where it gets things wrong and maintaining it when something changes. That is people's work, and it hasn't dropped by 75%.
We often see two traps. The first is sending whole documents when a fragment would do: past 100,000 tokens, Haiku 5.5 multiplies the price by five. The second is using the most expensive model for everything out of convenience. Most small business tasks are handled just as well by a small model, and the big one is kept for what really needs it. If you want to see how this connects with tools like n8n or Make, we explain it in n8n, Make or Zapier: which to choose.
What I would do this week
No grand plans. This is what I'd recommend to any small business already using AI or thinking about it:
- Dig out the list of what you ruled out as too expensive. If in 2025 you got a quote to automate something and it didn't add up, run the numbers again at these prices.
- If you already run an automation on Haiku 4.5 or a large model, test Haiku 5.5 alongside it with a hundred real cases and compare results before switching.
- Check the size of your requests. If any go over 100,000 tokens, ask yourself whether all that text needs to be sent.
- Don't marry one provider. Haiku 5.5 and GPT-6 Luna cost the same. Build the automation so that switching models means changing one line, not rebuilding the project.
- Set limits before granting access to tools. Cheap doesn't mean it can do anything. We cover it in what AI agents really automate.
What we think
We've been telling our clients for months that model prices would stop being the problem. This week it happened. What separates the company that gets value from AI from the one that doesn't is no longer the token budget, it's having someone who knows what to automate first, how to measure it and when to stop.
We also think it's healthy that Anthropic and OpenAI are competing on price at the same tier. For a small business, competition between providers is the best guarantee that nobody will suddenly hike your bill.
If you want us to look at it with you
We design and maintain AI agents and automations for small businesses, and if you need someone to make these technical decisions for you on an ongoing basis, we do it as an external CTO.
Book a thirty-minute call. We go through your team's repetitive tasks, what it would cost to automate them at today's prices and which one we'd start with. You leave with concrete numbers. No commitment.
Sources
This article draws on the following publications, consulted on October 9, 2026:
Frequently asked questions
$0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens. Above that size, $0.50 and $2.50. Reading cached content costs $0.01 per million.
90% cheaper for requests under 100,000 tokens and 50% for larger ones. Anthropic estimates an average saving of 75% on real workloads.
They have the same base price. Anthropic publishes results where Haiku 5.5 beats Luna on several tests, but those are the vendor's own figures. The sensible approach is to test both on a hundred real cases from your business and keep whichever handles your task best.
In tokens, very little: classifying a few thousand emails a month or extracting data from a thousand invoices costs cents or a few dollars a month with these models. The significant cost lies in designing the automation, integrating it with your systems and maintaining it.
Repetitive, high-volume ones: classifying emails or tickets, summarising conversations, extracting data from documents, answering frequent questions and tasks that a main agent hands off to smaller ones.
On the Claude Platform with the identifier claude-haiku-5-5, on Amazon Web Services, Google Cloud and Microsoft Azure, and also in Claude.ai on free and paid plans.
If a request goes over 100,000 tokens, Haiku 5.5 charges five times more per token. It's best to send only the part of the document you need or split it into chunks.