On 30 July, OpenAI cut the price of running its GPT-5.6 models through the API — the small Luna model by about 80 per cent, the mid-range Terra model by about 20 per cent. It reads like the sort of announcement that should show up on your next software bill. For most Australian small businesses it will not show up anywhere, because almost none of them buy AI in the unit that just got cheaper. That gap is worth understanding: it separates a genuine saving from a headline you can safely ignore.
There are two ways to buy AI, and they move independently
The first is by the seat. You pay a fixed amount per person per month — Microsoft 365 Copilot, a ChatGPT business plan, the AI features switched on inside your accounting or CRM package. The bill depends on how many people are licensed, not on how much they use it.
The second is by the token. A token is roughly three-quarters of a word. You are charged for the text going into the model and the text coming back out, quoted per million tokens, and you only see a bill like this if someone has wired the model up directly — a script, an integration, a tool a developer or agency built for you.
The 30 July cut was entirely on the second one. Nothing about it touches a per-seat subscription.
What the AI price cut actually changes
The new rates are US$0.20 per million input tokens and US$1.20 per million output tokens for Luna, and US$2 and US$12 for Terra. Sol, the largest model, did not get a price cut but is now reported to run about two and a half times faster in the API. The figures come from OpenAI’s own announcement on 30 July and were reported independently by eWeek and Forbes the following day.
Put those numbers against a real job. Say you want to pull the supplier, date, total and GST out of 2,000 invoices a month, and each one costs you roughly 1,500 tokens going in and 200 coming back. That is three million input tokens and 400,000 output tokens. On Luna, that is about US$1.08 a month. Before the cut the same work was about five times that, so a shade over US$5. On Terra, the more capable model, about US$10.80.
Those were not large numbers before, and they are trivial now. That is the point worth taking: for high-volume text work, the token price has stopped deciding whether an idea is viable. What decides it is whether the output is right often enough to trust, and what a wrong answer costs when it slips through. The cut removes cost as an objection. It does not remove accuracy as a requirement, and that was always the harder half.
What it does not change: the number on your invoice
Per-seat prices went the other way this year. On Microsoft’s Australian business pricing today, the Microsoft 365 Copilot Business add-on is AU$26.91 per user per month on a promotional rate, down from a list price of AU$31.40, and that sits on top of an underlying Microsoft 365 plan. Business Standard with Copilot included is AU$35.20 per user per month. For a ten-person office that is a little over AU$4,200 a year. The 30 July announcement moves none of it.
That is not vendors pocketing the difference out of spite. A seat price is not a resale of tokens: it bundles the interface, the security and compliance wrapper, support, and — the part businesses actually want — a bill that is the same every month regardless of use. None of those costs fall when the cost of serving a model falls. I would expect the two prices to keep drifting apart, and I would not budget on a token cut eventually reaching your subscription.
There is a second reason not to read too much into the headline percentage. Forbes made the point in its coverage that the visible per-token rate sits on top of the engineering, the checking and the rework, and those are what the bill is really made of. A model that is 80 per cent cheaper and wrong 5 per cent of the time is not cheap if someone has to review every record anyway.
Working out which one you are
A short exercise, worth doing once:
- Pull up the last two months of software invoices. If every AI line is charged per user per month, you buy seats, and this announcement is not about you.
- Look for anything metered — an OpenAI, Anthropic or Google API key on the company card, or an agency running something on your behalf. That is the bill that just fell, so ask for the new monthly figure.
- If you are on tokens, check which model name your integration actually calls. A cut on Luna does nothing at all if the code names an older model. Somebody has to change that line.
- If you have a quote for something that runs a model over every record, ask for the token cost to be broken out from the labour. Only one of those two just got cheaper, and it was never the big half.
What to do about it
If you buy AI by the seat, the honest answer is: nothing, and be sceptical of anyone who tells you it means you should be paying less. If you buy it by the token, spend half an hour on the new rate card, because a batch job that was marginal on cost a year ago probably is not any more.
The more useful version of the question is not what AI costs to run, but which of your repetitive data jobs would survive being automated at all — which is about how clean and consistent the underlying records are, not about the price per million tokens. That is the sort of thing a data health check is for, and it is the part that has not got cheaper.