The price of a million tokens has fallen roughly an order of magnitude in two years, and it is still falling. Every few months a provider announces that the same capability now costs a fraction of what it did, and the announcement is received as good news by everyone building on top.
It mostly is good news. It is also a strategy, and understanding whose strategy it is changes what you should build.
Prices are not falling only because the technology improved
Efficiency gains are real. Better serving infrastructure, quantisation, distillation, cheaper hardware per unit of throughput - all of that genuinely lowers the cost floor, and some of the reduction is simply passed through.
But the size and timing of the cuts do not track efficiency. They track each other. A provider cuts, a competitor matches within weeks, and the new price frequently sits below what independent estimates suggest inference actually costs at that scale. That is not a cost curve. That is a market being contested.
There are three distinct reasons a large company drives the price of something toward zero, and they lead to very different outcomes for you.
**To commoditise a complement.** If your revenue comes from cloud infrastructure, enterprise software or an advertising platform, cheap model access makes your real product more valuable. The model is not the business; it is the thing that makes the business bigger. Prices in this scenario go to near-zero and stay there, because nobody is trying to earn a margin on them.
**To buy distribution.** Below-cost pricing to establish default status is the oldest play in software. It ends when the market consolidates, and the price at that point depends entirely on how hard it has become to leave.
**To defend against being leapfrogged.** When capability differences between the leading models are small and shrinking, price becomes the only visible axis of competition. This produces the fastest cuts and the least stable ones.
The three look identical on a pricing page. They diverge sharply in year three.
What this means if you build on top of it
Three consequences, in order of how much they should affect your decisions today.
**Your margin is someone else's strategy.** If your product's economics depend on inference being cheap, you have taken a position on someone else's pricing decisions. That position has been profitable so far. It is still a position, and it is one you cannot hedge.
**Everything thin gets absorbed.** The clearest pattern of the past two years is that a product which wraps one capability in a nice interface becomes a feature of the underlying platform, usually within a year, usually announced with no warning. This is not hostility; it is the natural direction of a platform that wants to be more useful. If your entire product is a prompt and a form, the platform will ship it.
**Cheap inference changes which businesses are possible, and it also ends some.** Falling prices open categories that were uneconomic - anything requiring thousands of calls per user per month. They simultaneously destroy the businesses whose value was that the capability was expensive and scarce.
The durable positions
What survives a price collapse is what has never been about the model.
Proprietary data the platform cannot obtain. Integrations that took a year of unpleasant work with systems nobody enjoys touching. Workflow ownership, where you are not a step in someone's process but the place the process lives. Distribution and trust in a market where being known matters more than being clever. Regulatory position, where the barrier is an approval rather than a capability.
Notice that this list is identical to the list from every previous platform shift. The commoditisation of a capability does not eliminate advantage. It moves advantage to whatever the capability cannot supply.
What to actually do about it
**Design for portability from the start.** Keep prompts, model selection and provider clients behind an interface you own. This costs very little on day one and saves an entire quarter later, when a provider changes pricing, deprecates a model, or your best option becomes someone else's.
**Price on outcome, not on cost-plus.** If you bill against your token spend, your revenue falls whenever a provider cuts prices, which is a strange business to be in. Price against what the customer's alternative costs - the person, the agency, the software they use today.
**Assume the capability is free next year and ask what is left.** If the answer is nothing, the position is temporary regardless of current growth.
The part that does not get said
For most people building on these platforms, the price war is straightforwardly good. Cheap capability is the reason a great deal of currently viable work is viable.
The honest caveat is only that the current price is a decision, not a law, and decisions can be revisited once the contest ends. Building as though today's price is permanent is the same class of error as building as though today's model rankings are permanent. Neither has been stable for longer than a few months at a time.
What we do
We build AI agents and trading terminals, and we publish the scored record rather than the pitch. Every forecast is serialised, hashed and timestamped into a Bitcoin block before publication, so it cannot be adjusted once the outcome is known, and it is scored publicly with the misses on the same page as the hits.
Educational content - not financial advice.