Plain-English Briefing

AI just got 80% cheaper — but probably not for you

By Charlie Essex · 1 August 2026 · 5 minute read

On Thursday OpenAI cut the price of one of its models by 80%, three weeks after launching it. That is a real cut and a big one. It is also, for most small businesses, a cut to a price you have never paid and are never billed for — and it is worth understanding why before somebody uses it to sell you something.

Five briefings in a row about AI going wrong is enough. This one is about money, and it cuts against my own interests in a couple of places, which is rather the point of writing it.

What actually happened

On 30 July OpenAI reduced the price of two of its GPT-5.6 models. The cheapest one, Luna, fell by 80% — from $1 to $0.20 per million words of input, and from $6 to $1.20 per million words of output, in round terms. The middle model, Terra, came down 20%. The flagship, Sol, did not move at all.

Those are API prices. That is the wholesale rate a developer pays when a piece of software talks to the model directly — charged by the word, metered like electricity, billed to whoever wrote the code.

It is not the price of ChatGPT. It is not the price of Copilot. It is not the number on the invoice your bookkeeper files every month.

A small firm’s AI bill — and which part of it fell

What you are billed for

Per-seat licences — per user, per month−80%
AI features inside tools you already rent−80%
Setting it up, and the time it takes to run−80%
Metered model usage, billed by the word−80%

Most small firms never see this last line at all

The cut applies to the metered layer only — the wholesale rate a developer pays per word. If nobody has built you something that calls the model directly, that line does not appear on your bill, and nothing you pay for AI changed on Thursday.

Why it fell

OpenAI says it passed on efficiency savings from building the model. That is true and it is incomplete. DeepSeek launched a new model the same day. Cheap, capable models — a good number of them Chinese, most of them with freely published weights — have been eating the bottom of the market all year, and the new Luna price undercuts DeepSeek on input while the flagship, where OpenAI still has the field more or less to itself, did not move a penny.

That is a price war, not a gift. Which matters more than it sounds, because of what it tells you about the price.

Four things this actually means

If somebody built you a tool, go and check the bill. This is the genuinely good news and I do not want it buried under my own argument. If you have a quoting tool, a document reader, a chatbot on your website — anything a developer wired directly into a model — and it runs on the cheaper tiers, the metered part of that bill has fallen substantially, automatically, with nothing to do at your end. Ask whoever built it which model it uses and what last month cost. That is a five-minute email with a real number at the end of it.

If you buy seats, nothing changed. Per-user, per-month licences are priced against what a seat is worth to you, not against what the words cost to produce. Look at your renewal figure and then look at last month’s. If they match, you have your answer.

A price that can fall 80% in three weeks can do other things in three weeks. I am not predicting a rise — I have no idea, and anybody who tells you they do is selling. The point is narrower and it survives either way: this is a number set by somebody else, for their reasons, on their timetable, and you find out afterwards. Thursday was a pleasant version of that. It is still the same arrangement.

Cheaper models make owning cheaper too. The efficiency work that let OpenAI cut its rate is the same work happening in the open models — the ones that run on a box in your building. A model that needs less computing power to answer is a model that runs on less hardware. Falling costs are not an argument against owning. Over the last two years they have quietly been the thing that made it possible for a firm of twenty people at all.

The reason to keep your own data in your own building was never that tokens were expensive. Cheaper tokens do not touch that argument in either direction.

When renting is the right answer

Plainly: often. If a handful of people use AI a few times a week to tidy up an email or summarise a document, and none of what they paste is anything you would mind a stranger reading, then a per-seat subscription is the correct purchase and this week made the underlying economics of it better rather than worse. I would tell you the same on a call, and I would rather say it here than have you discover I only ever reach one conclusion.

Owning starts to make sense at a different point — when the AI needs to see the things you would not paste into a public box. Your prices. Your drawings. Your customer history. Your methods. At that point the question stops being what a million words cost and starts being where those documents end up, and no price cut moves that needle.

We have worked the comparison out properly over three years, including an honest section on when renting wins.

Worth doing this week

Find out which of the two you actually buy. Seats, metered usage, or both. Plenty of firms have quietly ended up with both and have never seen them side by side.
If you buy metered usage, ask for the new figure. The cut is automatic but the saving is invisible unless somebody looks.
Write down what your AI spend buys you. Not what it costs — what it does. If you cannot name the hours it saves, the price falling 80% has not helped you, because you were not getting anything for it at any price.

Our free AI policy pack includes a tools register, which is the short piece of admin that answers the first one. Personalised in about two minutes, no email, and nothing you type leaves your browser.

Get your free AI policy pack

Where this came from

More plain-English briefings on the Inferred blog. If you would rather talk it through, the fifteen-minute call is free and I will tell you if there is nothing to do.