Output tokens are the bill. Generation runs a full forward pass per token; prefill processes your whole prompt in parallel for roughly a tenth of the per-token energy. A short prompt with a long answer costs far more than a long prompt with a short one — and a reasoning model's hidden trace counts as output.
Same workload, every model class
Identical request volume and token shape, routed to different tiers. The jump to the reasoning tier is not incremental — it is the single largest controllable factor in an inference carbon budget.
| Model class | Wh / 1k output tok | kWh / mo | kg CO2e / mo | Power cost / mo |
|---|
Same workload, every region
Identical compute, identical energy — only the grid changes. This is the cheapest emissions reduction available to anyone whose provider lets them pick a region.
| Region | g CO2e / kWh | kg CO2e / mo | t CO2e / yr | vs your region |
|---|
Why the input/output split matters more than the token total
Almost every attempt to estimate AI energy multiplies total tokens by a single per-token number, and that is wrong in a way that matters. Transformer inference has two phases with completely different cost profiles. Prefill reads your entire prompt in one heavily parallel pass, so a 1,500-token prompt costs barely more than a 150-token one on modern serving stacks. Decode generates one token at a time, each requiring a full pass over the model weights, so output length scales energy close to linearly. The practical consequence is that stuffing more context into a prompt is comparatively cheap and letting a model ramble is expensive — the exact opposite of the intuition most teams start with, and the opposite of how the pricing sheet is shaped, since providers charge input at a discount too but nowhere near a 10× one.
The reasoning tier is where footprints go to die
A reasoning model's thinking trace is output. It runs through the decoder exactly like the visible answer does, it just never reaches the user. A request that produces 200 visible tokens after 2,000 tokens of deliberation has generated 2,200 tokens, and if that model is also an order of magnitude larger per token than the mid-tier alternative, the energy per request can be fifty times higher than the same task answered directly. Run the model-class table above with your own volume and the arithmetic is stark. This is also the most tractable lever: routing straightforward requests to a smaller tier and reserving reasoning for the cases that need it cuts both the bill and the footprint at once. Size that trade with the model routing savings calculator, and check what the thinking tokens are costing you in dollars with the reasoning token cost calculator.
Energy is a rounding error next to what you pay — and that is the point
Run the default numbers and the electricity behind two million mid-tier requests costs tens of dollars against an API bill in the thousands. Nobody switches provider to save that. The reason to compute it anyway is that the carbon number, unlike the energy cost, ends up in a CSRD or customer-questionnaire disclosure where it has to be defensible, and increasingly in procurement conversations where a supplier is asked for emissions per unit of service. The region row is the one to act on: identical compute in a low-carbon grid emits a fifteenth of what it does in a coal-heavy one, which no amount of model optimisation can match. If the disclosure obligation itself is what brought you here, price the surrounding governance programme with the EU AI Act compliance cost calculator.