AI API Pricing Just Fell. What That Changes for a Small Shop

Sometime in June you probably priced an idea and put it down. Rewrite every product description in the catalogue. Draft a first reply to every support email overnight. Summarise nine months of reviews into something you could act on. You looked at the API pricing, multiplied it by the number of items, and decided it was not worth the fuss for what it would cost.

That arithmetic changed on 30 July 2026. OpenAI cut the price of GPT-5.6 Luna, its cheapest tier, by 80 percent, and Terra, the middle one, by 20 percent. The flagship stayed where it was. Whether that changes anything for you depends entirely on which of those jobs you were actually pricing, and this piece works through the numbers rather than the announcement.

What actually changed on 30 July 2026?

Two of the three GPT-5.6 tiers got cheaper and the top one did not. According to OpenAI's published API pricing table, Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, with cached input at $0.02. Terra sits at $2 and $12. Sol, the flagship, is unchanged at $5 and $30.

The Decoder reported the previous Luna rates as $1 and $6, which is where the 80 percent figure comes from, and noted the company's own framing: a task that used to cost a dollar on last year's leading models now runs at roughly six cents. Unite.AI's write up adds the detail most coverage skipped, that the reductions carried through to the Batch and Flex tiers and to cached input, and that the three models had only reached general availability on 9 July. Three weeks between launch price and a four fifths cut is the part worth noticing.

The company also replaced its old Priority Processing option with something called Fast mode, which runs the flagship at up to two and a half times the speed for twice the price. So the same announcement made the cheap end cheaper and gave the expensive end a more expensive option. That is not a contradiction. It is a lineup being pulled apart at both ends, and it tells you where the volume is expected to come from.

What does a real shop job cost at the new prices?

Almost nothing, which is the honest answer and the one nobody puts in a headline. The table below prices five jobs a small merchant actually runs, using token counts from real prompt shapes: a product description takes roughly 600 tokens of input if you feed it the attributes and a paragraph of brand guidance, and returns about 250. A support reply needs more context and returns less. A review summary is mostly input.

Read the fourth column and then read the third. The entire annual social caption budget went from the price of a coffee to the price of a stamp. If you shelved one of these jobs in June because of what the model would cost, you shelved it for a reason that was already wrong, and this cut simply makes it more obviously wrong. The real bill for all five jobs together, at the new rates, is under five dollars.

The column that should worry you is the last one. Terra costs ten times Luna. Sol, unchanged, costs twenty five times it. A merchant who wires an integration to the flagship out of caution, then runs it across a catalogue, is not spending $0.42 on those descriptions. They are spending eleven dollars, and at ten thousand items, a hundred and ten. Model choice, not model price, is what moves a small AI bill, which is the argument we made at length in our comparison of what a real feature costs across the main APIs.

Why did the price fall three weeks after launch?

OpenAI's stated reason is efficiency and the market's reason is competition, and both are probably true. The company credits kernel level work that cut serving costs by 20 percent, plus a redesigned speculative decoding setup worth more than 15 percent on token generation throughput. Those are genuine engineering gains and they do show up as headroom on a price card.

The other half is visible in the timing. The Decoder put the cut in the context of low cost Chinese providers and Microsoft pushing its own cheaper models, and framed it as OpenAI adopting the pricing posture of its competitors rather than defending a premium. Three weeks is not how long it takes to discover an efficiency. It is how long it takes to read the usage numbers after a launch.

What did not get cheaper?

Three things, and each one is a way for a bill to grow while the price per token falls. The first is long context. On the same pricing page, every GPT-5.6 tier charges a higher rate once a request crosses into the long context band: input doubles and output rises by half. Luna at $0.20 becomes Luna at $0.40, which is still cheap, but the multiplier applies to the whole request rather than the overflow.

The second is Fast mode. Two and a half times the speed for two times the price is a fair trade when a customer is waiting on a page, and a poor one for anything running in a queue overnight. Nothing in a catalogue rewrite needs it.

The third is the flagship, which sat out this round entirely. If your integration was built when the middle tier was the sensible default, the cut has quietly widened the gap between what you use and what you could use. That is worth one afternoon of checking.

Is caching now the thing that matters most?

For a shop with a catalogue, yes. Cached input on Luna is $0.02 per million tokens against $0.20 for fresh input, so context you send repeatedly costs a tenth as much once the cache is warm. The pattern that benefits is exactly the one a merchant has: the same brand guidelines, the same tone rules, the same shipping policy, sent in front of every one of two thousand different products.

Concretely, a review summarisation job that ships 8,000 tokens of shared context per product pays about $0.0016 for that context uncached and about $0.00016 cached. Across five hundred products the difference is under a dollar, which sounds like nothing until you run the same job weekly for a year. The habit matters more than the sum: whoever builds your integration should be putting stable instructions at the front of the prompt and the variable product data at the back, because that ordering is what makes a cache hit possible at all. We went through the mechanics of this in the piece on why the token bill lands higher than the estimate.

Does cheaper inference change what a small shop should do?

Less than the headline suggests, and the evidence for that is public. The US Census Bureau's Business Trends and Outlook Survey release covering 14 December 2025 to 3 May 2026 found overall business AI use hovering between 17 and 20 percent, with retail trade at 14 percent current use and 17 percent expected within six months. Firms with fewer than twenty employees showed no significant change across the period and sat under 20 percent, while firms with 250 or more employees reached 37 percent.

Model prices fell across that entire window. Adoption among the smallest firms did not move. If price were the binding constraint, those two lines would not look like that. What actually stops a merchant is narrower and less glamorous: nobody has an hour to check two thousand generated descriptions, nobody is sure whether publishing them is allowed, and the first attempt produced copy that sounded like every other shop.

That last problem is the expensive one. A generated description costs you 0.042 cents and fifteen seconds of your attention. At two thousand items the model time is under a dollar and the attention is eight hours. Cutting the model price by 80 percent moves the total cost of that project by roughly nothing, because the money was never in the tokens.

How does this compare with what else is on the market?

At the cheap end it undercuts everything, and at the middle it makes the middle crowded. Unite.AI's analysis put Luna's new rates at roughly four to five times below Anthropic's Haiku 4.5 on both input and output, and noted that Terra at $2 and $12 now sits under the $3 and $15 rates Claude Sonnet 5 is due to charge from 31 August 2026. Those are the comparisons that matter to a buyer, and they are also the comparisons that expire fastest.

Which is the argument for not building your shop around one vendor's price card. The gap between the cheapest usable model and the second cheapest has been under a factor of two for most of this year, and it keeps swapping hands. Anyone whose integration reads the model name from configuration can take the next cut for free. Anyone who hard coded it will read another announcement like this one and do nothing about it, because changing it means opening the code again.

How do you find your own token counts?

Run twenty real jobs and read the usage field in the response, because every estimate in this article is a shape and yours will differ. The number that surprises people is input, not output. A prompt carrying your full brand guide, a product's attributes, three example descriptions and a policy note gets to 600 tokens quickly, and if you paste the whole style guide instead of the relevant paragraph it gets to 6,000 without you noticing.

Output is easier to reason about because you specified it. Two hundred and fifty tokens is roughly 180 words of English, which is a normal product description. If you asked for 180 words and the model returns 400, you are paying for the overrun on every single item, and the fix is a sentence in the prompt rather than a cheaper tier. Measuring once at the start is what turns a per token price into a number you can actually plan against, and it is the step most people skip straight past on their way to a spreadsheet full of guesses.

Which jobs move from too expensive to worth trying?

Honestly, very few, because the tiers were already affordable. The jobs that genuinely change are the ones where volume is enormous and each individual result is low stakes. Scoring every incoming review for sentiment and routing the angry ones. Tagging a photo library. Classifying support emails into buckets before a human opens them. Producing a first pass alt text for ten thousand images.

Those share a shape. Each call is cheap, the volume is large enough that the old price added up, a wrong answer is recoverable, and no customer sees the raw output before you do. That is where an 80 percent cut compounds into a real number.

What has not changed is anything customer facing and singular. Your homepage copy, your refund policy, the reply to the customer who is already annoyed. Those were never blocked by cost, and running them on the cheapest tier to save four cents is the wrong economy. If you want the practical version of that split, we set it out in the guide to where AI earns its keep in a one person business.

How do you stop the bill creeping back?

Set a hard spend cap on the API key before you start, not after the first surprise. Every provider offers one and almost nobody switches it on until they have a reason to. A runaway retry loop on the cheapest tier can still produce a bill you did not plan for, and the cheaper the tokens are, the less anybody watches.

Then pick the tier per job rather than per project. Classification and tagging go to the cheap tier. Anything a customer reads unedited goes up a level. Nothing goes to Fast mode unless a person is genuinely waiting. And keep the prompt structure stable so the cache does its work.

Last, count the hours. The only line in an AI project that scales badly for a small business is human review, and the way to shrink it is to run the job on a sample of fifty, read all fifty, and fix the prompt before touching the other two thousand. That discipline saves more than any price cut announced this year. If you are weighing what an AI build costs when you add the platform on top of the tokens, our own pricing page shows the credit model we use, and it is priced off real inference cost for the same reason.

What this means

Cheap inference at the low end is now a commodity and the vendors have stopped pretending otherwise. Luna at $0.20 in and $1.20 out is priced to be embedded in software rather than sold as intelligence, and the flagship holding at $5 and $30 tells you where the margin is meant to live. Expect that gap to keep widening.

For a merchant the practical reading is dull and useful. Stop treating model cost as a reason to postpone. Start treating the review hour as the real budget line, because it is the one that did not fall by 80 percent and never will. Choose the cheapest tier that clears your quality bar for that specific job, cache what repeats, cap the spend, and check a sample before you publish anything with your name on it.

And do not rebuild anything this week. The next cut is already coming, from whichever vendor reads its usage numbers next, and the shops that win are the ones whose integration does not care which model is behind it.