Token prices fell 80%. Your agent bill didn't.

Here's a fact that should make agent costs go down and instead makes them go up: over the last two years, the price of a token has fallen roughly 80%. Frontier-class capability now costs a fraction of what it did. By every intuition, your AI bill should be shrinking.

Ask any team running agents in production whether their bill is shrinking. It isn't. It's the fastest-growing line item they have. That's the agentic cost paradox, and it's worth understanding precisely, because it's the reason cost-cutting projects keep failing.

Cheaper tokens, more tokens, way more tokens

A chatbot makes one model call per message. An agent makes many — it plans, calls tools, reads the results, reflects, retries, and loops until it's done. A single agent task can be dozens of calls, each dragging the full context along with it.

So when the price per token drops 80% but the number of tokens per task goes up 10x — and then you run that task hundreds of thousands of times a month — the unit price is irrelevant. You optimized the one number that stopped mattering.

Why "just use a cheaper model" doesn't fix it

The obvious move is to route everything to a cheaper model. Teams try it, quality drops on the hard tasks, they revert, and conclude cost optimization is a tradeoff against quality. It isn't — that's just what happens when routing is a global switch instead of a per-run decision.

Most agent runs are easy and a small model handles them perfectly. A minority are hard and need a frontier model. The waste isn't the frontier model — it's sending the easy runs to it because you have no way to tell them apart at request time. Fix that, and cost drops with no quality loss. Miss it, and every optimization is a blunt instrument.

It's a control problem, not a pricing problem

Step back and the paradox dissolves. Token price is a market number you don't control. Tokens-per-task, calls-per-run, retries, context bloat, and model-per-request are all operational numbers you absolutely control — if you can see them. Most teams can't. They see a provider invoice, aggregated and after the fact.

That's why cost programs stall: they attack the price (which fell and didn't help) instead of the operations (which are where the money actually goes). Put a cost on every run, a budget on every agent, and a routing policy that matches each call to the cheapest model that can do it — and the bill finally tracks the token price down instead of decoupling from it.

The 80% drop is real. You're just not allowed to feel it until you control the runs.

Cheaper tokens were never going to save you. Operating your agents will.