Environmental footprint
The AI that counts its CO2 and water.
Spine reserves big models for big jobs, leaves calculations to code, and shows you the footprint of your usage every month.
In Spine
Your footprint, spelled out every month.
On the Usage screen, next to the credits.
- The current month, projected
- Marked “≈” until the end of the cycle.
- Hosting and usage
- The fixed share, then each model level.
- Everyone's share
- Pro rata to each person's credits.
- Concrete comparisons
- Kilometers by car, meals, showers.
≈ 12 kgof CO2e per month
5 kg so far, Oct 28- 83 km in a gas or diesel car
- 53 km on a short-haul flight
- 2 days of gas heating (100 m² home)
- 2 meals with beef
≈ 43 Lof water per month
19 L so far, Oct 28- 7 toilet flushes
- 4 minutes with the faucet running
- 1/2 of a 5-minute shower
- 1/5 of a car wash with a hose
The right model
No big model for a small question.
The larger a model, the more computation each answer takes. By default, Spine answers with the fastest and least expensive model in your plan.
-
EverydayFind, write, reply.default
-
AnalysisCompare, summarize, advise.
-
In-depthTenders, hundreds of pages: the most powerful.
Titles, summaries, background tasks: always at the “Everyday” level.
A small, well-informed model beats a large one that guesses.
- An organized memory
- Typed, linked documents: the model receives a well-ordered brief.
- The right search
- Only the useful passages are handed to it.
- Your methods
- Written once in skills, applied every time.
Orders of magnitude
About ten times less energy for everyday tasks.
The energy used by 1,000 answers of the same length, according to EcoLogits.
- Everyday
- In-depth
The figures and the models
| Model family | Level | Model assessed | Energy for 1,000 answers, midpoint | Range (Wh) |
|---|---|---|---|---|
| OpenAI | Everyday | gpt-6-luna | 70 Wh | 67 to 72 |
| In-depth | gpt-5.6-terra, similar to gpt-6.1-terra | 1,000 Wh | 750 to 1,300 | |
| Anthropic | Everyday | claude-haiku-5-5 | 53 Wh | 29 to 77 |
| In-depth | claude-sonnet-5-5 | 750 Wh | 580 to 910 | |
| Mistral AI | Everyday | mistral-small-2603 | 120 Wh | single value |
| In-depth | mistral-large-4 | 1,600 Wh | single value | |
| Everyday | gemini-3.5-flash-lite | 130 Wh | 35 to 230 | |
| In-depth | gemini-3.1-pro-preview, similar to gemini-4-argon | 6,600 Wh | 4,200 to 9,000 |
EcoLogits 0.11.3, October 28, 2026: 500 output tokens per answer, midpoint of the range, estimated size for closed models. The open-weight models served by Scaleway have no figure of their own in EcoLogits.
Calculations
For calculations, code. Not AI.
A language model doesn't do the math: it writes an answer. Spine has the calculation written once, in code, then runs it exactly the same way every time.
Calculators written and verified by the assistant, business apps built for you, automations that call them: the calculation is done in code.
- Reproducible. Same values, same result.
- Traceable. The code, its version, and the values are kept.
- No AI model. The server's processor is enough.
The method
Where the numbers come from.
Public sources, at every step.
-
Consumption
Scaleway for hosting, the model's API for each call.
-
Conversion
Scaleway, and EcoLogits, an open-source library cited by ADEME (the French Agency for Ecological Transition): CO2e and water, based on the electricity mix of the data center's country.
-
Comparisons
ADEME, Cieau (the French water information center), and INC (France's National Consumer Institute), to turn the figures into kilometers or showers.
Version 1.0, dated October 28, 2026.
Hosting
A transparent hosting provider.
Since September 2026, every new Spine instance has been running in France, hosted by Scaleway, which publishes the footprint of its products, project by project.
- By project and by product
- CO2e and water, in the console and via API.
- Manufacturing included
- Servers, network, and buildings, following ADEME's methodological framework.
- The actual grid mix
- With no deduction for renewable electricity contracts.
6 times less CO2efor the same AI computation done in France rather than in the United States, where EcoLogits places OpenAI, Anthropic, and Google. Spine's line-up of open-weight models runs in Paris.EcoLogits 0.11.3, small Mistral model, manufacturing included, October 28, 2026.
Our stance
Numbers, not adjectives.
No AI is impact-free. We don't call it “green”: we tell you what it weighs.
What we provide
- Orders of magnitude, rounded
- The boundary, including what isn't counted
- A published, versioned method
- A number for your carbon accounting
What we won't say
- “Green AI,” “carbon-neutral”
- Offsetting passed off as reduction
- A number with no date or method
- A total that hides its limits
Digital technology has a weight. Spine shows you yours.
Carbon accounting and CSR
A real number for your carbon accounting.
The footprint of your usage, with its method, rather than a ratio applied to an invoice.
In practice
What ADEME recommends. What Spine does.
Four recommendations from its opinion on generative AI, published in July 2026.
-
ADEMEAsk whether generative AI is necessary.
SpineCalculations are done in code, not with an AI model.
-
ADEMEUse models suited to the need, and make that choice visible.
SpineThe level is chosen with each message: “Everyday” by default, the most powerful for big jobs.
-
ADEMEBe transparent about the method.
SpineA published method, with its sources, its boundary, and its limits.
-
ADEMEDon't stop at the figure for a single request.
SpineThe footprint of a month of usage, for the company and for each person.