The post at minimaxir.com this month carried a headline that doubles as a question: “OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?” It pulled 346 points and 335 comments on Hacker News, which is the internet’s way of saying the number landed. Everyone has an answer, and the answers are the same three: inference got cheaper, the models got more efficient, and two well-funded rivals are buying market share from each other.
All three are probably true. None of them explains the thing that actually matters about the cut, which is that it is not principally aimed at winning usage. It is aimed at ending a negotiation.
The Number Measures a Price List, Not a Cost Curve
“Up to 80%” is a phrase from a marketing document, not an audited cost disclosure. Somewhere there is a specific model at a specific context length on a specific tier that got five times cheaper, and somewhere else there is a frontier product that didn’t get cheaper at all. This is how price lists work in every industry that has ever had one: the discount lives on the SKU you’re trying to move, and the thing everybody actually wants holds its number.
That isn’t a scandal. It’s just how the arithmetic behind a headline gets flattened into a trend. The 335-comment thread under the post spends most of its energy on the cost-curve explanation — GPUs, kernels, distillation, the usual.
Fine. But cost is the least interesting input into a price. If falling cost set prices, no software company would ever have charged per seat.
Why Nobody Buys at a Falling Price
Here is the problem the labs created for themselves and are now paying to fix. From roughly the moment enterprise AI got a budget code, the price per token has been falling fast enough that the rational move for any buyer was to wait. Sign a twelve-month commitment at today’s rate and you have pre-paid for the privilege of watching your competitor sign the same deal at half the price in April. Every CFO knows this instinct from cloud contracts, and every CFO has been burned by it.
A market with a visibly falling price is a market where the pilot never graduates. Nobody builds production on a number that hasn’t stopped moving.
Which means a cut of this size is not primarily a customer-acquisition tool. It’s a floor-building exercise. You collapse the list price toward something that looks stable and defensible so the number stops being a variable in the customer’s model and becomes a constant. Boring is the point. A boring cost line is what a budget committee can approve.
The bears read an 80% cut as margin surrender, and therefore as evidence the business is weaker than advertised. The bulls read it as adoption. Both are staring at the same number and missing that its real function is administrative: OpenAI and Anthropic are spending margin to convert a speculative purchase into a recurring one. That isn’t distress. It also isn’t generosity. It’s two companies deciding that predictability is worth more than the last 80 points of gross margin, because predictability is what a purchase order requires.
The Margin Went to the Fortune 500
And here is the part that should bother people who normally cheer for price competition. The discount does not go to a teenager building a side project. It goes to whoever can move volume — and volume lives inside organizations with procurement departments, vendor scorecards, master service agreements, and a person whose entire job is extracting the last fifteen percent.
“A year ago we were begging for a pilot,” one procurement manager for a Midwestern insurer said on a vendor-selection call, reading the words off a rubric in front of her. “Now we have two of them bidding against each other and a third one that keeps calling.”
That is the actual mechanism. The price war is a transfer. Two venture-funded, hyperscaler-subsidized vendors are moving margin off their own balance sheets and onto the P&L of the largest companies in America — companies that did not ask for a subsidy, would never describe it as one, and are constitutionally opposed to industrial policy when anyone else runs it.
The list price is what a small business in Ohio pays. The 80% is what a Fortune 500 procurement group negotiates after the third meeting. That gap is the whole story, and it is invisible in the headline because a headline is a single number.
The uncomfortable corollary, for people who assume markets sort these things out: the constraint on AI adoption from here is not capability, and it isn’t cost either. It’s a signature. Whoever makes an eighteen-month enterprise budget cycle move faster than the model release cadence wins the decade — and that person is not a researcher, and probably not an engineer.
So yes, an 80% cut in a few months is remarkable. Just be clear about what got cheaper. Not intelligence — intelligence was always going to get cheaper. What got cheaper was the argument you have to make to your own finance department. Someone has decided to make it for you, and sent the bill to a balance sheet you don’t own.