OpenAI made its cheapest model 80% cheaper: what that means for you

If you build anything with AI — an app, an internal tool, an automation — the price of the model is usually the number that decides what's viable. On 30 July 2026, one of those numbers fell by 80% overnight.

OpenAI sells GPT-5.6 in three tiers. Luna is the fast, cheap one; Terra is the everyday middle; Sol is the most capable. Luna's price dropped from $1 to $0.20 per million input tokens, and from $6 to $1.20 for output — an 80% cut. Terra fell 20%. Sol's price did not change, though it gained an option to run 2.5 times faster for twice the cost.

A million tokens is roughly 750,000 words. At the new price, feeding a model a million words of input costs about twenty pence. That is the level at which whole categories of idea — summarise every support ticket, check every document, watch every log — stop being expensive experiments and start being ordinary features.

There is a detail here that affects people who never touch an API. OpenAI said the lower prices are also reflected in how usage is counted against paid subscriptions in Codex and ChatGPT Work. In plain terms: the same monthly plan now buys you far more work on the cheaper models before you hit a limit.

Two reasons, and both are true. The first is efficiency, and OpenAI's version of it is unusual: the company says it used GPT-5.6 itself, working inside its Codex coding tool, to study its own live traffic and rewrite parts of the low-level software that runs the model on graphics chips. It reports that work cut serving costs by around 20% and improved generation efficiency by a further 15%.

The second reason is competition, which the announcement does not mention. Three weeks earlier the GPT-5.6 family had launched at the old prices. In between, Anthropic released Claude Opus 5, Google shipped a cheaper Gemini, and the Chinese lab DeepSeek was selling a capable model at roughly a seventh of Luna's old price. When a rival charges fourteen cents for something you charge a dollar for, the price tends to find its way down.

For you, the reason matters less than the effect. AI capability per pound has been falling steeply all year, and this was one of the sharper steps. The practical habit worth forming: re-check the pricing page of whatever you use every few months, because the assumptions you built your budget on are probably out of date.

A useful mental model, now that the gap between tiers is so wide. The cheap tier is for volume: sorting, tagging, summarising, drafting, answering the straightforward half of your questions. The middle tier is for work where a mistake costs you something but not much. The expensive tier is for the jobs where getting it right matters more than what it costs — complex reasoning, long multi-step tasks, anything you would otherwise have paid a professional for.

Most people and most products default to the expensive tier for everything, which is comfortable and wasteful. With Luna at twenty pence a million words, the more sensible default is the reverse: start cheap, measure whether the output is good enough for the specific job, and move up only where it visibly isn't. That single habit will usually cut an AI bill by more than any negotiation ever would.

It is also worth knowing that this is not a one-off. The cost of a given level of AI capability has fallen repeatedly through 2026, driven by a mix of genuine engineering progress and fierce competition between American, Chinese and open-weight providers. If something you wanted to build was too expensive six months ago, it is worth pricing again.