What a token costs you and what it burns are unrelated numbers
Vals AI found a model fifteen times cheaper per token with the same footprint, days before its chief executive said token spend may start to eclipse payroll.
What you pay for a token tells you almost nothing about what the token costs to produce.
Vals AI published an audit on 3 September of what open-weight models actually consume, and that is the finding underneath it. DeepSeek V4 Pro is nearly fifteen times cheaper per output token than Kimi K3 and carries roughly the same environmental footprint. Per completed task it is about six times cheaper.
Same work. Same physical cost. Wildly different invoice.
The bars span two orders of magnitude for the same set of tasks, which is the whole argument in one picture.
How they got there, and why only Chinese models appear
Vals took token usage from its own index of 2,157 long-horizon agentic tasks across finance, coding, healthcare and law, then pushed it through EcoLogits, an open-source estimator that needs a model's parameter counts to work (no counts, no estimate).
Which is why the sample is what it is. Fourteen of the sixteen models are Chinese, one comes from Thinking Machines Lab and one from Mistral. Nobody can run this exercise on GPT or Claude, because you need the parameter counts and those are the whole secret.
The arithmetic nobody in the report did
Kimi K3 needs about 1.7 million watt-hours for the full index. Spread across the task count, that is roughly 788 watt-hours per task.
Sam Altman and Google both put an average text query between 0.24 and 0.34 watt-hours last year. So one agentic task on the most accurate open model costs something like 2,600 average chatbot queries.
At the other end, Ling 3.0 Flash comes in near 11 watt-hours a task. About seventy times less, for a model built to answer simple things quickly.
That is the shape of the problem. Capability is being bought in orders of magnitude, and Bloomberg's write-up of the same report puts the gap between a quick question and building a web app at up to ten thousand times.
So why does price not track any of this?
Because price is set by market share, and Vals says so plainly. Providers offer tokens at steep discounts while trying to capture share, which distorts the relationship between what you pay and what you use.
And cheap tokens can be less efficient tokens. A model that thinks longer bills you less per token and more per job, which is exactly the trap a procurement spreadsheet is built to fall into.
Then the chief executive said the quiet part
Six days later, Rayan Krishnan told a podcast that "token spend may start to eclipse salary spend."
That sentence turns an environmental report into a finance one. If inference is heading toward the size of payroll, then a per-token price is a payroll system that bills by the keystroke. You would not run a headcount plan that way (imagine a staffing budget denominated in emails sent).
Krishnan's other line is the practical one. The ability of a company "to make its evals legible in order to solve this ROI calculus is going to be the reason why that company wins out over its competitors in the long term."
The unit of procurement has to move
Our expectation is that it moves from tokens to completed tasks, and that the vendors who benefit are the ones whose models finish quickly rather than the ones with the lowest sticker price. We would look for at least one major enterprise AI contract reported on a per-task basis within twelve months. If the market is still quoting dollars per million tokens in late 2027 with no efficiency term attached, we were early or simply wrong.
The honest caveat is that these are estimates on assumed hardware, not meter readings, and Vals counts several of the labs it measured among its customers. It says so itself. Take the ratios seriously and the absolute numbers loosely (the ranking is the useful part).
One question to put to your own vendor, then. How many watt-hours did my last job take, and if you cannot answer that, what exactly are you charging me for?
