On 30 July 2026, OpenAI cut the price of two of the three models in the GPT-5.6 family. The cheapest of them, Luna, became 80% cheaper. The models had reached general availability on 9 July — meaning the price fell three weeks after launch.
The news was read everywhere as "AI is getting cheaper". It is, but not in the way the headline suggests — and the difference changes business decisions.
The Exact Numbers
- GPT-5.6 Luna: input from US$ 1.00 to US$ 0.20 and output from US$ 6.00 to US$ 1.20 per million tokens. An 80% cut on both.
- GPT-5.6 Terra: input from US$ 2.50 to US$ 2.00 and output from US$ 15.00 to US$ 12.00. A 20% cut.
- GPT-5.6 Sol: nothing changed — still US$ 5.00 input and US$ 30.00 output, the same price as the previous generation.
What Almost Nobody Reported: the Top Did Not Move
This is the part that vanishes from the headlines. The most capable model did not get a cent cheaper. What fell was the bottom and the middle of the range.
That is not a detail, it is the shape of the market: intelligence that has become a commodity collapses in price, while frontier intelligence holds its value. The practical consequence for a company is direct — the repetitive, well-defined task became drastically cheaper; the hard task did not.
If your use case is classifying messages, extracting data from documents, answering frequent questions or summarising text, your cost just dropped to a fifth. If your case demands long reasoning and delicate judgement, your cost is exactly what it was three weeks ago.
Why It Fell — and Why It Will Keep Falling
OpenAI attributed the reduction to infrastructure optimisations: rewriting GPU kernels, which cut roughly 20% of the cost of serving the models, and improvements to speculative decoding, which raised token-generation efficiency by more than 15%.
Note that none of this is about the model getting smarter. It is execution engineering. And execution engineering improves continuously — it happened with hosting, with storage and with bandwidth. The price of running AI will keep falling, and you do not need to believe a forecast to act accordingly: just look at the history of any digital input.
There is competitive pressure too. At the new price, Luna is roughly five times cheaper than Anthropic's Haiku 4.5 on input, and Terra now costs less than Claude Sonnet 5's introductory pricing. Sol, meanwhile, remains more expensive than Opus 5 on output. In other words: the fight is being fought at the bottom, not at the top.
Three Decisions This Changes in Your Company
1. Stop waiting for it to get cheaper. It is the most common reaction and the most expensive. If the price falls every month, there will always be a reason to postpone — and the cost of waiting is not the discount you missed, it is the process that stayed manual the whole time. Whoever started in July at the old price already knows what to automate; whoever waited will start from zero, cheaper, months behind.
2. Do not marry a model. A price that moves 80% in three weeks is by definition unstable. What protects you is not choosing correctly today, but building so you can switch tomorrow without redoing everything. That is an architecture decision: nearly free if considered at the start, far too expensive if remembered later.
3. Distrust long contracts quoted at today's price. A vendor locking you in for one or two years on the current token price is freezing your margin against a curve that only goes down. An AI contract in 2026 calls for a price review clause, not loyalty.
What Gets Expensive When the Token Gets Cheap
We are an agency that implements AI, so let us say the thing that does not serve the easy pitch: the model's cost is almost never the bottleneck of a project. It was already small before the cut and it is smaller now.
What costs is everything else: mapping the process worth automating, wiring the AI into the systems the company already runs, cleaning up data that is a mess, defining what happens when the answer comes out wrong, and training the people who will live with it. None of that got 80% cheaper today. None of it will.
When the input becomes a commodity, the advantage moves to whoever knows where to apply it. It is the same conclusion we reached analysing open models and doing the real math on AI in business — different roads, same destination.
Frequently Asked Questions
Did the ChatGPT I use get cheaper?
No. The change is to API pricing, charged per token, used by those integrating AI into systems. Consumer monthly subscriptions are unchanged.
Should I switch to the cheapest model now?
Only if it can handle your task. A cheap model that gets things wrong costs more than an expensive one that gets them right — the cost of rework and errors rarely makes it onto the spreadsheet, and usually exceeds the price gap.
Is it worth waiting for the next price cut?
No. There will always be a next one. The gain from automating a process now is recurring and compounding; a future discount applies only to what you consume afterwards.
How do I know if my case fits the cheap model?
By the nature of the task. Repetitive, with a verifiable answer and clear criteria: almost always yes. Ambiguous, with long context and serious consequences when wrong: usually not.