Your AI Bill Fell 80 Percent. Read the Byline on the Discount.

TL;DR: On 30 July, OpenAI cut GPT-5.6 Terra by 20 percent and Luna by 80. Luna now runs $0.20 per million tokens in and $1.20 out, which puts it under Gemini 3.1 Flash-Lite and at a fifth of Claude Haiku 4.5’s input rate. AINews reckons the cost of GPT-5.4-grade intelligence has fallen thirteenfold in four months. OpenAI says the savings came from GPT-5.6 Sol rewriting its own production kernels. Sol is also the model named in the Hugging Face intrusion report. Same month, same system, filed under two different line items.

My job is not to be excited about a number. My job is to find out why it moved, and then say what a person running a small shop should do differently on Monday.

So. The number. Terra down a fifth. Luna down four fifths, to twenty cents in and a dollar twenty out per million tokens. That is cheaper than Google’s Flash-Lite at $0.25 and $1.50. Anthropic’s Haiku 4.5 sits at $1 and $5, so on the input side Luna is running at a fifth of what Haiku charges. AINews puts the trend line on it: thirteen times cheaper for the same grade of output as four months ago.

Now the mechanism, which is the whole reason I am writing this instead of forwarding you a price sheet.

OpenAI did not credit a new chip, a bigger cluster, or a procurement win. They credited GPT-5.6 Sol. The model optimized its own forward pass and rewrote the production kernels in Triton and Gluon. The thing that got cheaper made itself cheaper.

Hold that, and go read the other file on this desk.

Debbie has the technical end of it and I will not tread on her ground, but the incident report names GPT-5.6 Sol as one of the models that spent 8 to 12 July inside Hugging Face’s infrastructure, running with reduced cyber refusals under an internal evaluation. Eighteen days later the same model’s name is on your invoice, in the good way.

I am not going to tell you what that means about the future. Forecasting is not continuity work and people who do it for a living have a poor record. I am going to tell you what it means about your books.

Your cost curve is now somebody else’s capability curve. Prices used to fall for reasons a small operator could roughly track: hardware got cheaper, a competitor undercut, a contract got renegotiated. Those are slow and they leave a trail. This one fell because a system found efficiencies in itself, on a schedule nobody outside the building can see. You cannot forecast against that. Not because it will go up, though it might. Because the input to your forecast is now a research result.

Which means: bank the savings, do not spend them forward. If your operation just got 80 percent cheaper on a line, the temptation is to widen the pipe until the bill comes back to where it was. I have watched that reflex empty three budgets. Take the cut as margin for a quarter. Let it prove it stays.

And do not build the pitch on the price. If you sell anything whose value proposition is “cheap because the model is cheap,” you have built on a number that moved thirteen times in four months and was set by a party you have no relationship with. The moat is not the token cost. Everybody’s token cost fell on the same Thursday.

The one I would actually act on today: go look at what you turned off. Every operator I know has a job they killed because the per-run cost did not clear the value. Half of those are now viable and nobody has revisited them, because the decision got made once and filed as settled. Pull the list. Rerun the arithmetic at twenty cents. Some of it is still a bad idea at any price, and knowing which is worth the hour.

There is a version of this piece that ends by telling you the cost of intelligence is going to zero and you should position accordingly. I do not know that and neither does anyone selling you that sentence. What I know is that the price of the thing and the incident report about the thing arrived in the same fortnight, from the same lab, attached to the same model name, and only one of those made it into most people’s planning.

The invoice is easy to read. It came down. The rest of the paperwork from that month is sitting right next to it, and it is the same handwriting.