Why your AI bill rises while token prices fall
Quick answer
For two years the question was what AI costs. That question is mostly answered: frontier prices are 84.5 percent below March 2023 and the small tiers fell about 96 percent. The next constraint is not price per token. It is how much work a person can review, trust and integrate in a day.
The number that changed
When a language model costs $45 per million tokens, you ration it. You write shorter prompts, you cache, you batch, you argue about seats. At $0.10 per million, rationing stops being the interesting problem.
That is where the small-model tier sits in October 2026: GPT-6 Luna and Claude Haiku 5.5 both at $0.10 input and $0.50 output per million tokens. A thousand support replies cost less than a coffee.
So the scarce input stops being tokens and becomes attention.
Three things that get expensive instead
Review capacity. If an agent drafts 400 documents a day and a person can read 40, you have not automated the work, you have automated the queue. The human step is now the throughput limit, and it costs far more per unit than the model does.
Integration debt. Cheap tokens make it tempting to bolt a model onto every workflow. Each one needs an owner, a fallback path and a log. Ten half-owned automations cost more to maintain than one good one, and none of that shows up on an API invoice.
Trust. A wrong answer from a free tool is an annoyance. A wrong answer from an agent that writes to your CRM is an incident. As inference gets cheaper, the cost shifts from generating output to verifying it.
What this looks like inside a company
The teams that benefit most from the price collapse are not the ones generating the most text. They are the ones who built a cheap review path: templates, diffs, sampling, and a clear rule for what a human must see before it leaves the building.
Teams without that path hit a ceiling fast. Volume goes up, quality drifts, and the first serious incident freezes the whole programme for a quarter.
The 2027 version of the question
If tokens keep falling 60 to 70 percent a year, the cost of a first draft approaches zero. The value moves to three places: knowing which draft to send, catching the one that is wrong, and connecting the output to a system that acts on it.
None of those are model problems. They are process problems, and they are priced in hours, not tokens.
One habit worth building now
Measure cost per finished task, including the human minute it takes to approve it. When the model line approaches zero, that second number is the only one left on the page.
Read next: how to measure cost per task, and the power crunch behind inference pricing.
Questions people ask
If token prices are falling, why is my AI bill going up?
Because usage grows faster than prices fall. Frontier prices are 84.5 percent below March 2023, but agents read hundreds of thousands of tokens per task, so total spend rises even as the unit price drops.
What replaces tokens as the constraint when prices drop?
Review capacity. If an agent drafts 400 documents a day and one person can check 40, the human step becomes the throughput limit, and it costs far more per unit than the model does.
Should I still track cost per token?
Track cost per full task instead, including the human minute it takes to approve the output. That second number is what stays visible when the model line approaches zero.