The 80 Percent Price Cut Shows Where AI's Real Value Sits Now

Share

Everyone pieces together the same strange picture this week. OpenAI and Anthropic both cut the price of their flagship models, in some cases by 80 percent over four months. For the past couple of years the working assumption in the industry was the opposite trend. Building frontier models kept getting more expensive, so most people expected token prices to stay high and drift down slowly. Instead the two biggest labs in the world slashed prices at a pace nobody predicted and both at once.

I think the price cuts matter much less for what they do to your API bill than for what they reveal about the business. An 80 percent cut in four months is not a promotion and it is not a sale. It is a signal, made in the most expensive way these companies know how, about what they believe they can charge for. When you slash the price of your most advanced model that fast, you are effectively admitting that raw model access is no longer where the durable value sits.

The numbers come from a developer's documented timeline that made the rounds last week. OpenAI cut the price of one flagship model by 80 percent to roughly 20 cents per million input tokens. Anthropic trimmed Claude pricing by up to 80 percent for certain tiers, and over the same stretch its top Opus model moved from about five dollars per million input tokens to four. No single cut is unusual in isolation. The combined speed is. A cache-heavy coding agent that switched from one flagship model to a newer one saw its effective cost drop roughly 75 percent, according to the same analysis.

Here is the part that seems strange on its face. The obvious reading is that the AI boom is deflating and the laboratories are scrambling for volume. I think there is a better reading, and it has less to do with weakness than with a market finally pricing itself honestly. Open weight models have gotten good enough for a huge share of real work, and they cost a fraction as much to run. The labs are not cutting prices primarily to win. They are cutting prices so they do not lose the commodity end of the market to models that undercut them on cost by design.

At the same time, it would be wrong to read this as value destruction. Cheaper tokens grow the market. When the marginal cost of intelligence falls that sharply, developers build applications they could not justify six months ago, voice agents that were too expensive, background processes that were not worth the tokens. Reported token spend has been roughly flat for two months even as prices fall, which is a remarkable thing in itself. It means usage is expanding fast enough to absorb the cheaper prices. An 80 percent price cut that leaves total spend unchanged is not a shrinking market. It is a rapidly expanding one being repriced underneath everyone who thought they were selling tokens.

There is a useful historical parallel in the early cloud years and the software price wars that preceded them. When big players cut prices in a market that is commoditizing, the ones who keep pricing power are rarely the ones with the cheapest atoms. They are the ones who control the relationship with the customer. The consumer internet taught the same lesson over and over. Value migrates to whoever owns distribution and the customer relationship, not to whoever happens to run the cheapest server. I keep coming back to that pattern because it is the lens through which this price war starts to make sense.

If token prices keep falling toward the floor set by open weight models, the premium that remains has to live downstream of the model itself. It lives in the workflow that is wired into how a company actually operates. It lives in the data a product has accumulated about its users and their business. It lives in the trust a customer has in an agent that has quietly done the job well for a year. It lives in the distribution that puts a model in front of a buyer at all. None of that gets cheaper because tokens got cheaper. That is the part I think founders should be watching closely.

For everyone building on top of these models, the situation cuts both ways. On one side, cheap frontier access is a genuine gift. The cost of experimenting and prototyping falls to almost nothing, and entire categories of application become economically sensible for the first time. On the other side, it removes a whole layer of accidental protection that a lot of startups have been leaning on without realizing it. A surprising amount of the AI software economy has quietly been reselling the raw capability of a frontier model and marking it up with a thin wrapper. When the base price collapses, so does the value of the wrapper. If most of a product's value is borrowed from the model it calls, then the model owner can always undercut you and reclaim that margin.

This is the irony that lands last. The laboratories slash their prices, and in doing so they invite, maybe force, the layer above them to compete on something they cannot easily be undercut on. If an application's only moat is that it happens to sit on a slightly newer model, then an 80 percent price cut at the base just as easily becomes a squeeze on the margin of the layer above. The value the application adds has to live somewhere the model provider does not and cannot go. It has to live close to the customer, in the form of work done and trust earned, not in the form of access to a smarter model.

Of course, none of this means every layer above the models is doomed, any more than cheap web hosting doomed every software company. Plenty of durable businesses were built on infrastructure that ultimately became commodities. The ones that survived were the ones that added something the infrastructure owner had no reason or ability to add, a domain-specific workflow, a trusted relationship, a distribution channel, a body of proprietary data. Cheap tokens make that distinction clearer rather than harder. They strip away the accidental value and leave the real value more visible.

I suspect the frontier premium is not disappearing altogether. It is concentrating. The price cuts push the bulk of usage toward cheap tiers and open weights, while the genuinely best models stay expensive relative to their performance edge. The result could be a market that looks like two worlds living side by side. In one, good enough intelligence is everywhere and almost free. In the other, a shrinking number of buyers pay a real premium for the top of the frontier. The danger for the laboratories, and the opening for everyone else, is that the gap between good enough and frontier keeps narrowing until only a small fraction of buyers ever need the very top.

So I do not celebrate the price war, and I do not fear it. It is the market behaving like a market, repricing something that turned out to be far cheaper and far more abundant than the early years of the gold rush suggested. The money is already moving, from the model itself to the layer around it that a user cannot live without. Companies that build their moat there, in workflow, data, distribution, and trust, will do fine as prices fall. The ones that were really just reselling access will feel the squeeze first.

Intelligence got cheaper by 80 percent in about four months. The things that keep a customer around, the trust, the workflows their habits have grown into, the data only you hold, did not get cheaper at all. That gap is where the real value now sits, and it was never in the model to begin with.