Cost and footprint calculator

Describe your model mix, the impact shows up as you type.

Public methodology
Usage profile

No context estimated: only output tokens are counted.

Your mixhow your traffic actually splits across models
Your mix rows
%

Total: 100%

Examples

Scenario result

CO₂e / month

2.4 kg

445 g to 7.1 kg

Cost / month

€4

That is about 17 km in a combustion car and 1 shower for this monthly volume.

Confidence is low on the CO₂ factor of Claude Haiku: interpret the figures with caution.

Energy / month

6.3 kWh

Water / month

37.6 L

Water: 2% for cooling, the rest to generate the electricity.

Tokens processed on this profile

  • Cache read0%
  • Cache write0%
  • Input0%
  • Output100%

One email with your numbers. Then at most 3 follow-ups over 15 days, only if your spend is large enough to measure, each with an unsubscribe link. No newsletter.

These numbers are estimates, at the providers' list prices. If they look really wrong to you, write to gaetan@tokenclimate.com.

“Output only” profile: only output tokens are counted. In agentic use, context and cache make up most of the volume: pick a profile to estimate them. The dashboard measures this ratio on your organisation's volumes, pulled every day.

52 models covered, each with its sheet and confidence level: all models.

Track these numbers every month, without typing them

The dashboard pulls these volumes on its own from your Anthropic, OpenAI or OpenRouter admin keys, and splits them by team and by person.

The demo dashboard: emissions with their range, budget, CO₂e per session and per million tokens, 30-day trend and split by team.
Fictional organisation, sample data.

Frequently asked questions

What exactly does the calculator count?
The four token classes billed by providers: input, cache write, cache read and output. The scope is that of the TokenClimate methodology: third-party datacenter, inference only, location-based by default. Every energy factor is sourced and versioned, and the full methodology is public.
Why aren't output tokens enough?
Because in agentic use they are only a small part of what the machine processes. On the coding agent profile, about 91% of the tokens processed are cache reads. A cache read costs about 0.08 times the energy of an input token (see the methodology's cache energy section): little per token, but the volume makes it a significant term of the total.
Where do the usage profile ratios come from?
Each profile is a sourced, versioned set of reference volumes, shown on screen with its assumption. The chat profile relies on the median of a conversational production trace (Splitwise, ISCA 2024). The coding agent profile is copied from the aggregates TraceLab measured on 2,676 real Claude Code sessions, 37 developers from a single lab. The real ratio varies from one organisation to another: the dashboard measures it on your real data.
I don't know my token volume, what now?
Nobody knows it by heart, which is what the simplified mode is for. It asks three questions, what you delegate to the tool, how many people use it and how often, then derives a monthly volume you can correct in the field. Tokens per session (96,900,000 / 2,676, TraceLab) and per exchange (150, Splitwise) come from published sources; the cadence is not guessed, you enter it. To place yourself, the methodology publishes the one TraceLab measured: 1.9 session per developer per week, or 8.0 per month. The result is a ballpark figure rounded to two significant digits, not a measurement.
Can a scenario be shared or cited?
Yes. The URL encodes the models, their volumes and the usage profile, nothing else. The figures are recomputed on every open with the current version of the methodology, server-side for emails. A cited scenario therefore stays verifiable and recomputable.