The footprint of a prompt
Research and methodology for the carbon comparison calculator.
Reviewed November 3, 2026 · Default AI measurement: May 2025.
What the default measures
The default is a historical Gemini Apps text-request scenario: 0.24 Wh, 0.26 mL of cooling water, and 0.03 gCO₂e per median request in May 2025, published by Google in August 2025. It is not a current average across AI models or a demonstrated lower bound.
Google includes active accelerators, host systems, idle capacity, and facility overhead. Carbon uses market-based electricity accounting and includes hardware embodied emissions. Water covers data-center cooling consumption. Training, end-user devices, and electricity-generation water are outside these figures. A median describes a typical request; multiplying it by many requests does not establish their actual total footprint.
Sources: Google — Measuring the Environmental Impact of Delivering AI at Google Scale (August 2025).
The heavy scenario is an assumption
The second preset scales all three Gemini metrics by 40: 9.6 Wh, 10.4 mL, and 1.2 gCO₂e. It illustrates a heavier request under the same assumed carbon and water intensities. It is not a measured category of professional use, a coding-agent task, or a proven conservative estimate.
Epoch's February 2025 model estimates about 0.3 Wh for GPT-4o with an assumed 500 output tokens, rising to about 2.5 Wh with 10,000 input tokens and 40 Wh with 100,000 input tokens. These depend on deployment and input-processing assumptions.
Oviedo et al.'s June 2026 revision models a median of 0.31 Wh per typical frontier-scale query (interquartile range 0.16–0.60 Wh), and 3.91 Wh (2.15–7.05 Wh) for a scenario with 15 times longer test-time scaling. Those intervals describe that study's model, not uncertainty bounds for every AI request. Neither energy estimate supplies a corresponding universal carbon or water factor.
Sources: Epoch AI — How much energy does ChatGPT use? (February 2025); Oviedo et al. — Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling (June 2026 revision).
Why the token total is illustrative
The calculator multiplies equivalent requests by an editable assumed output length: 500 tokens for Gemini or 8,000 for the heavy scenario. The 500-token assumption comes from Epoch's GPT-4o model, not Google's measured Gemini request length. Neither value establishes a measured carbon cost per token.
Input processing, caching, hidden reasoning, model size, hardware, and serving utilization change the relationship between tokens and energy. A September 2026 Greenpixie methodology explicitly separates input and output tokens and models uncertainty for proprietary deployments. The earlier claim on this page that per-token research did not exist was too broad.
Sources: Epoch AI — How much energy does ChatGPT use? (February 2025); Greenpixie — AI Token Methodology (September 2026 preprint).
Activity factors and their boundaries
Driving: 400 g CO₂ per mile, EPA's approximate passenger-vehicle tailpipe average. Fuel production and vehicle manufacturing are excluded.
Winter heating: 26.5 kg CO₂ per day, calculated as an assumed 5 therms/day × EPA's 5.3 kg CO₂/therm combustion factor. Five therms is an illustrative household assumption, not a verified winter average.
Chicago–Austin flight: 200 kg per passenger, one way. This retained illustrative estimate has not been checked against a route-specific model. Aircraft, cabin, load factor, and aviation's non-CO₂ effects are not modeled.
Office day: an unsourced illustrative 15 kg per worker-day for building energy and commuting.
Quarter-pound burger: an unsourced illustrative 3 kg CO₂e per burger; production boundaries were not independently verified.
Coffee: an unsourced illustrative 90 g CO₂ per cup for brewing energy only, excluding agriculture, processing, and transport.
Each activity factor can be edited. Changing the activity restores that activity's default factor and amount. Reset restores the selected AI preset and activity factor while keeping the chosen amount.
Sources: US EPA — Greenhouse Gas Emissions from a Typical Passenger Vehicle; US EPA — Greenhouse Gas Equivalencies Calculator: Calculations and References.
How to interpret the comparison
Equivalent requests = activity amount × grams of carbon emissions per activity unit ÷ grams CO₂e per AI request. Assumed output tokens = equivalent requests × assumed output tokens per request. Energy and water totals describe only the AI side; they do not imply equal energy or water use for the activity.
Outputs are rounded to two significant digits. This is display rounding, not a statistical confidence interval. The activity side mixes direct CO₂ and CO₂e estimates and different lifecycle boundaries, so comparisons are approximate rather than harmonized lifecycle assessments. Editing carbon independently of energy or water creates a custom scenario; the calculator does not infer a new grid intensity.
Provider-specific figures should not be swapped without checking their scope. For example, Mistral's July 2025 disclosure gives 1.14 gCO₂e and 45 mL water for a 400-token Le Chat response, with a different methodology. It cannot directly replace Google's figures or rank the services on its own.
The flight, office, burger, and coffee defaults remain evidence gaps. Replace them with a documented factor matching your actual activity before using this page for a consequential comparison.
Sources: Mistral AI — Our contribution to a global environmental standard for AI (July 2025).
What changed in this review
Reviewed November 3, 2026 against the linked primary sources. The Google measurement date is May 2025; the original site report was compiled June 14, 2026. The calculator now identifies historical measurements and illustrative assumptions, removes the unsupported lower-bound and conservative labels, makes activity factors editable, and rejects invalid input instead of silently replacing it with defaults. The heavy water factor is now exactly 40 × 0.26 = 10.4 mL.
This revision replaces the previous verification-vote narrative with explicit source attribution and limitations. It also removes narrow-method values that were inconsistent with the cited paper's Table 2. No new environmental measurements were performed for this site.
Sources
- Google — Measuring the Environmental Impact of Delivering AI at Google Scale (August 2025)
- Epoch AI — How much energy does ChatGPT use? (February 2025)
- Oviedo et al. — Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling (June 2026 revision)
- Greenpixie — AI Token Methodology (September 2026 preprint)
- US EPA — Greenhouse Gas Emissions from a Typical Passenger Vehicle
- US EPA — Greenhouse Gas Equivalencies Calculator: Calculations and References
- Mistral AI — Our contribution to a global environmental standard for AI (July 2025)