By Brad Ferris · 10 October 2026
On 7 October Anthropic released Claude Haiku 5.5, its smallest production model, priced at US$0.10 per million input tokens and US$0.50 per million output tokens. Anthropic puts that at roughly 75% below the previous Haiku on average, and it halved the price of cached reads on its mid-tier Sonnet model in the same release. In the same week McKinsey published The new economics of AI, which makes an argument that sits awkwardly next to the price cut: token prices keep falling, and total AI spend inside organisations keeps rising.
Both things are true at once, and the gap between them is where a lot of mid-market AI budgets are about to go wrong.
When something useful gets cheaper, people use more of it. Economists call this the Jevons effect, and AI is a clean example. A process that failed its business case six months ago on inference cost now passes. A team that rationed its use of a model stops rationing. An agent that used to be given one attempt at a task is now given five, because each attempt costs less.
McKinsey's response is to tell CEOs to run AI consumption the way mature businesses run cloud spend: with forecasting, demand management and a FinOps discipline, using cost per token and energy per token as the defining measures. That is sensible advice for a large enterprise with a cloud cost team. Most Australian businesses with 50 to 300 staff do not have one. Their AI spend sits across a handful of software subscriptions, a usage-billed API account someone set up on a credit card, and whatever the staff are paying for themselves.
The second problem is quieter. A token is a unit of text, and the price list is written in tokens. The business pays for completed work.
An independent read of the new model, reported by Latent Space using Artificial Analysis data, shows why those two can diverge. On Artificial Analysis's testing, Haiku 5.5 used roughly 162,000 output tokens per task, against roughly 50,000 for the OpenAI model it was being compared with. A model that is cheaper per token and three times wordier per task can still come out ahead, but the headline price stops telling you whether it does. The same testing recorded a 40% hallucination rate on Artificial Analysis's knowledge benchmark, which matters if the job involves looking things up and stating them as fact.
None of this makes the model a poor choice. Anthropic itself says its larger models remain better for complex agentic coding, and positions Haiku as the cost-efficient option for high-volume tasks. The point is that "75% cheaper" is a statement about a price list. Whether your invoice processing, customer triage or quote preparation gets 75% cheaper depends on how many tokens each run consumes, how often it has to be retried, and how much human checking sits behind it.
1. Cost per completed task. Pick the two or three processes where AI is doing real volume and work out what one finished unit costs: one invoice coded, one enquiry answered, one report drafted. Include retries and the time someone spends checking the output. This is the number to compare across models and vendors, and it is the only one that tells you whether a cheaper model is cheaper for you.
2. Volume trend. Look at usage over the last three months, by process where you can. If volume is rising faster than unit price is falling, the line item will grow even as every vendor announces price cuts. That is a healthy sign if the extra volume is doing useful work. It is a warning if nobody can say what the extra volume is for.
3. The ceiling. Developer and writer Simon Willison argued this month that we are going to need default hard budget caps on pretty much everything. His concern is agents: software that can call tools, loop and retry without a person approving each step. Once an agent can act on its own, a spending limit stops being a billing preference and becomes a control, in the same category as a payment approval limit. Every usage-billed account the business holds should have a hard monthly cap set at a level someone has chosen on purpose.
The practical consequence of the price cut is positive. Work that was too expensive to automate at the start of the year may now be worth automating, and any business case rejected on inference cost in the first half of 2026 deserves a second look with the new numbers.
The risk is in how the saving gets booked. If the board or the owner hears "AI just got 75% cheaper" and reads it as a budget reduction, the business will be surprised twice: once when usage expands into the headroom, and again when a verbose model or an unattended agent turns a low unit price into a high monthly total.
The fix is ordinary management accounting applied to a new cost line. Know what a finished job costs, know which way volume is heading, and put a ceiling on anything that can spend money without asking. A mid-sized business does not need a FinOps team to do that. It needs one person who owns the AI line in the accounts and looks at it monthly, the same way someone already looks at freight, fuel or software licences.
Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.