The AI price war nobody's calling a price war
GPT-5.6 Luna dropped 80%. Gemini crossed a billion users. The cost of frontier AI is falling faster than at any point since it launched — here's what's real.
Stated by the maker, publisher, or platform holder.
Something is happening to the price of AI that isn't getting talked about as the story it is.
In the space of about a month, OpenAI cut one of its models by 80%, Google pushed out deliberately cheap models that forced the response, and OpenAI crossed a billion active users right after dropping its prices. The cost of running frontier-class AI is falling faster than at any point since these models existed.
That's a bigger deal than most of the model-launch headlines it's buried under, because price is the thing that decides what actually gets built. Here's what's confirmed, and what it means.
The numbers, confirmed
These come from the companies' own pricing pages and announcements, so they're solid — though with the caveat that AI prices now change so fast that any figure is a snapshot.
The pattern underneath: OpenAI's cut came about a week and a half after Google introduced its low-cost Gemini 3.6 Flash and Flash-Lite. One lab drops a cheap model, the other responds within days. That's not coincidence — that's a price war, even though nobody's using the phrase.
Why prices are falling
Two forces, and they pull in the same direction.
The first is competition, and it's not just the American labs. OpenAI's own framing when it cut prices named the pressure directly: enterprises are less willing to deploy expensive models without a clear return, and Chinese startups plus Google and Microsoft have been undercutting on cost. DeepSeek's models sit at a fraction of Western flagship prices. When a good-enough model exists at a tenth of the cost, the expensive one has to justify the gap or drop.
The second is that the labs have worked out something about their own business: volume beats margin. OpenAI cutting Luna 80% and then announcing a billion users in the same breath isn't two stories. Cheap models bring users, users bring lock-in, lock-in is worth more long-term than a high per-token price. Altman's stated framing is that each generation should advance capability and efficiency — which is the polite way of saying prices go down.
The part that's genuinely underappreciated
The flagship tier has converged to near-identical pricing.
The top models from the major labs now sit clustered together — several around $5 input and $25–30 output per million tokens. When the prices are that close, the per-token cost stops being the thing you choose on. What decides it is which model is actually better at your specific task, which is a much harder question than reading a price sheet.
That convergence is the real signal. It means the "which AI is cheapest" question, which dominated enterprise decisions for two years, is quietly dying. The models that win from here win on quality per task, not price per token — because the price per token is becoming the same everywhere.
What it means for you
If you use these tools through a subscription, this mostly reaches you as "the free tier got much better," which it has — cheaper models mean the free defaults get more capable.
If you build on the APIs, the ground is moving under you monthly, and the right move is to stop optimising for today's cheapest model and start being able to switch. The prices are converging; the winner keeps changing.
And the bigger picture: AI getting dramatically cheaper is what determines which uses become viable. Things that made no economic sense at last year's prices make sense now. That's where the next wave of actual products comes from — not from a smarter model, but from a cheap-enough one.
The prices will have moved again by the time you read this. That's the story.