On 30 July, OpenAI cut the price of GPT-5.6 Luna by 80 per cent. Luna, the cheapest and fastest model in its current line-up, now runs at about $0.20 per million input tokens and $1.20 per million output tokens. The mid tier, Terra, came down 20 per cent. The flagship, Sol, did not move. This is not a rounding error. It is the price of machine thinking falling through the floor, and it changes what is worth automating.
A word on the jargon. A “token” is a chunk of text, very roughly three-quarters of a word. “Input tokens” are what you send the model, “output tokens” are what it writes back. Agent-style tools rack up huge token counts because they make many small calls to finish one job, so when the per-token price drops, tools that were too dear to run suddenly pay for themselves.
Why a solo operator should care
Jobs that would have cost real money a year ago are now close to free at volume. Luna is built for exactly the high-volume grunt work that eats your afternoons: sorting inbound enquiries, classifying documents, tagging support tickets, triaging a messy inbox, pulling structured data out of unstructured mess. It can use tools and follow multi-step instructions, so it is a genuine building block, not just a chatbot.
The reason it is happening is worth naming out loud. The frontier labs are competing on price now because cheaper rivals have put them under pressure, and OpenAI says its own efficiency gains (including the model helping rewrite its serving code) let it pass the savings down. When giants undercut each other, the small operator with an API key is the one who pockets the difference.
What to do this week
Run an audit. List the repetitive, high-volume tasks you or your team still do by hand: the sorting, the labelling, the first-draft summarising. Move the boring, well-defined ones to a cheap model like Luna, and keep an expensive flagship for the hard 10 per cent that needs real judgement. A good first project is a triage job. Point a cheap model at your inbound (email, form fills, DMs), have it categorise each item and draft a first response, and leave the sending to you.
Two cautions. Cheap is not clever. Luna is for clearly defined, repeatable tasks, not the judgement calls, so do not hand it your pricing strategy. And watch the output side. Output tokens cost several times more than input, so a job that writes long answers can still add up. Measure a small batch, check the quality, then scale.
The headline number is the 80 per cent. The real story is that “let a model handle it” is now a sensible default for the dull, repetitive half of the job, at a price that no longer needs a spreadsheet to justify.
Did you know: output tokens on Luna are priced at roughly six times the input rate, which is why the running cost of a chatty agent is driven less by what you ask and more by how much it says back.
Related on Top Tool Stack: Inkling-Small: own your model · DeepSeek’s free V4 Flash