Claude Haiku 5.5 at $0.10 per million tokens: which jobs an SMB can hand to it instead of a pricier model

On 7 October Anthropic put on sale a model that costs $0.10 per million input tokens and $0.50 per million output tokens ¹. If you have a chatbot on your website or an automation that writes product descriptions, this is the price that concerns you.

The short answer: for short, repetitive jobs it is worth trying this week. On quality, nobody can give you a figure that holds for your business. You measure it yourself, on real replies, before you switch off the model you use today.

What it really costs, item by item

The previous Haiku, version 4.5, cost $1 per million input tokens and $5 per million output tokens ². The new Haiku 5.5 starts at $0.10 and $0.50 ¹. On paper that is an enormous cut.

There is a threshold, though. Simon Willison, who follows model prices closely, points out that above that threshold the price goes up by 5 times ².

Then there is the price rise you do not see on the price list. The new model uses a less generous tokenizer: according to Willison's measurements, the same long prompt uses about 1.25 times the tokens it used with Haiku 4.5 ². So do not take today's bill and divide it by the new price. Redo the sums on your real tokens.

The tricky part is prompt length

The model accepts up to 1 million tokens of context ¹. That does not mean it is worth filling it.

A customer's question, with your instructions and a few pages of FAQs, usually stays well below the threshold. The same goes for a product description generated from one line of a catalogue. Here you really do get the low price.

If instead you feed the model whole contracts, complete catalogues or long conversations, the story changes. OpenAI's GPT-6 Luna has the same price as Haiku up to 100,000 tokens, and rises only to $0.20 and $0.75 beyond 272,000 ². Willison puts it bluntly: above 100,000 tokens Luna looks like the better deal ².

Reasoning cannot be switched off, but it can be lowered

The new Haiku always reasons before replying: you cannot disable it, and the default level is medium ². For a chatbot this matters, because reasoning costs time and tokens.

In Willison's test, at the lowest level one reply cost 0.0936 cents and arrived in 7 seconds ². At the highest level the same request took 5 minutes and 9 seconds, at 3.3826 cents ². No customer waits five minutes in a chat.

My advice: for standard replies and product descriptions, set the low level, and raise it only if your tests show errors.

Which jobs to move and which not

I would move first the jobs where a mistake costs little and the volume is high. Replies to recurring questions about shipping, opening hours and returns. Product descriptions written from data you already have. Sorting incoming emails.

I would not move straight away the replies to delicate complaints, quotes, and anything with legal weight. There the saving is small compared with the damage of a wrong reply.

On quality, the only data point you have is that Haiku reports higher benchmark scores than Luna ². But benchmarks know nothing about your customers or your catalogue. They count for little until you have run your own test.

If you have a Max or Team subscription, check the settings

Along with the model, Anthropic added monthly API credits to its subscriptions: $100 a month for Max 5x, $200 for Max 20x, up to $500 for Team plans, shared between users ². You turn them on from Settings, under billing ².

Credits do not carry over from one month to the next ². If you have them, they are the simplest way to run tests without paying anything extra. You can also switch off automatic top-up, so that requests stop when the balance runs out ². Do it: it is your insurance against a surprise invoice.

What to do on Monday

Take a sample of real questions that reached your chatbot in the last month, with the replies the current model gave.

Have the new Haiku answer the same questions, at the low reasoning level. Then have the two sets read by whoever looks after your customers, without telling them which is which.

Check how long your typical prompt is. If it often goes over the threshold, compare Luna as well before deciding.

If you spend only a few euros a month on AI today, you can also do nothing. The time you spend migrating is worth more than what you save.

If you would rather have help setting up the comparison, AIR does this with SMBs. But it is a test you can run on your own in an afternoon, and the afternoon pays for itself with the first invoice.