Our methodology

What we measure

We measure the operational emissions of the inference phase: the electricity a datacenter consumes to answer your requests.

These exclusions are consistent with Watershed, Climatiq, and most enterprise Scope 2 accounting practices. We state this scope explicitly so you can incorporate it correctly into CSRD reporting.

The calculation

• tokens — input and output tokens are counted separately, since output generation is markedly more energy-intensive per token than reading input.
• J / 1,000 tokens — the energy (in joules) a model consumes per 1,000 tokens processed, distinct for input vs. output.
• PUE — Power Usage Effectiveness of the datacenter running the model (total facility power ÷ compute power).
• gCO₂e / kWh — carbon intensity of the electricity grid powering that datacenter.
• ÷ 3,600,000 — converts joules to kilowatt-hours (1 kWh = 3,600,000 J).

Input tokens are processed in parallel (prefill) — markedly less energy-intensive than output tokens, which are generated sequentially, one at a time. Not distinguishing the two understates the true carbon cost of long responses.

Confidence levels — what they actually mean

We show a confidence level for every model. It's not a rating of the model itself — it's our estimate of the uncertainty in the carbon number we calculate for it.

high

Official provider data with a published methodology, plus independent third-party verification of that exact model. Typical uncertainty: ±10-20%.

medium

A peer-reviewed academic benchmark measured against real APIs, or an official provider disclosure. Typical uncertainty: ±20-50%.

low

No infrastructure-level benchmark exists for this model — estimated by analogy to a measured model's architecture or size class. Typical uncertainty: ±50-70%.

No model is currently “high confidence.” No AI lab publishes per-token inference energy with a verifiable, real-time methodology. That gap is exactly why a page like this one needs to exist.

On Gemini specifically: Google disclosed in May 2025 that its median query consumes 0.24 Wh and emits 0.03 gCO₂. That second figure uses market-based factors that fold in Google's PPAs — it reflects Google's clean-energy commitments, not the physical carbon intensity of the grid it drew from. On a location-based basis, our calculation gives roughly 0.09 gCO₂ for the same median query.

Reasoning models — a different ballgame

“Reasoning” models (o3, o1, DeepSeek-R1, Claude Extended Thinking) generate a long internal chain of thought before answering. That multiplies output tokens — and the energy consumed — well beyond what a standard model uses for the same visible response.

Emissio detects reasoning models automatically and applies their distinct factors — tracking an o3 request as if it were GPT-4o would understate its footprint by roughly 17x. A short GPT-4o query costs about 0.34 Wh; the same prompt to o3 costs about 7.03 Wh.

Scope and exclusions

We don't calculate emissions from training. GPT-4-class models are estimated to have emitted on the order of hundreds of tonnes of CO₂e during training — but that cost is amortized across billions of requests and isn't triggered by your usage. We focus on what your AI consumption decisions actually affect: inference.

Similarly, your own hardware (laptop, internal servers) and network connection aren't included — they're shared with everything else you do and aren't specific to AI usage.

Frequently asked questions

Are your numbers certified?

No — no third party currently certifies AI inference carbon calculations. We apply the same methodology as the most recent peer-reviewed academic papers and publish our sources so they can be independently verified.

How do you handle models with no published data?

We estimate them by analogy to a measured model of similar architecture and size, and mark them “low confidence” with roughly ±60% uncertainty rather than presenting a false precision.

Why don't you use Google's official CO₂ figure for Gemini?

We do use it for energy (0.24 Wh per median query). For carbon, we prefer location-based factors over market-based factors that include PPAs, consistent with the GHG Protocol's Scope 2 guidance.

How often do you update these factors?

Whenever a major academic benchmark or official provider disclosure is published. Last updated: September 2026.