On Wednesday, OpenAI cut the price of one of its main models by 80 percent. Not a sale. Not a promo code. The list price — slashed, three weeks after the model launched.
GPT-5.6 Luna, the fast workhorse tier of OpenAI's current lineup, dropped from $1 to 20 cents per million input tokens. The mid-tier Terra model got a 20% trim. The flagship, Sol, didn't move at all. CNBC's read on it was blunt: business customers are getting cost-sensitive, and the labs are feeling it. Anthropic ran a similar move on its own mid-tier earlier this summer. The competition among AI labs is quietly shifting from "who has the smartest model" to "who has the best economics."
You will probably never buy a token in your life. So why should a business owner care? Because tokens are the raw ingredient of every AI tool you rent. The chatbot on your website, the AI features inside your email platform, the meeting-notes tool, the writing assistant baked into your CRM — all of them buy tokens wholesale from labs like OpenAI and sell them back to you inside a subscription. When the wholesale price of the ingredient falls 80% overnight, the economics of everything built on top of it move too. Here's what that means for you, practically.
The "AI add-on" fee is now mostly margin
A lot of software vendors charge $20–$50 per seat per month for their "AI features." Some of that pays for real engineering. But the biggest line item underneath it — the model usage itself — just got dramatically cheaper for tasks that run on workhorse-tier models, which is most of what these features do: drafting, summarizing, classifying, autocomplete.
That doesn't mean your vendor will cut your bill next week. It means the pressure now runs in your favor. Expect AI features that were premium add-ons to start showing up in base plans, because some competitor will use the cheaper tokens to do exactly that. If you're paying a chunky AI upcharge, it's a fair moment to ask your vendor what it's actually buying you.
The model didn't get smarter this week. It got cheaper. For a business, that's the same news.
Don't sign an annual AI contract right now
This is the second time in a month a major lab has cut prices, and the cut came three weeks after launch. That's not a market that has found its floor. Any AI tool pricing you're quoted today was built on last quarter's token costs. If a vendor is pushing hard for an annual commitment on an AI product, the discount for committing needs to be large enough to beat the direction prices are moving on their own. Month-to-month keeps you on the right side of that curve.
Recheck your "too expensive to automate" list
This is the part most owners will miss. Somewhere in the last year, you probably looked at an automation and decided it didn't pencil out — classifying every inbound email, summarizing every customer call, drafting a reply to every review, scoring every lead. Per-item AI costs made high-volume jobs feel indulgent.
Run the math again. At Luna's new price, a task that reads about a thousand tokens — a longish email — costs two hundredths of a cent. You could run every email your business received this year through a classify-and-draft step for less than the cost of one coffee. The jobs that were borderline in January are rounding errors in August. The constraint on automating the repetitive parts of your business is no longer price. It's whether you've set it up.
One honest caveat: the flagship model's price didn't move, and that's not an accident. Hard reasoning still commands a premium, and the labs know it. The practical skill here is matching the model to the job — cheap workhorse models for volume tasks, the expensive one for the judgment calls. That split is where the savings actually live.
The headline this week was about OpenAI's pricing strategy. The story underneath it is simpler: the raw material of automation keeps getting cheaper, faster than almost anyone budgets for. The owners who win aren't the ones watching the price charts. They're the ones who already have the pipes laid, so every price drop flows straight to them.