Every published figure we know of for the energy or carbon cost of a single AI query, with what it includes and where it comes from. 2 of the 5 are official provider disclosures; the rest are executive statements or independent research. Typical text queries land between about 0.2 and 0.35 Wh.
Published figures, newest first
Microsoft Research
Independent research · 2026Typical chatbot query on production-optimized, frontier-scale serving (median 300 output tokens), and a long reasoning query (about 5,000 output tokens)
- 0.31 Wh median per typical query (interquartile range 0.16–0.60 Wh)
- 3.91 Wh median for a long reasoning query — about 13× a typical one
Scope: Modelled estimate for optimized deployments, published in the journal Joule; not a measurement of one company's production fleet.
Figures as summarized by Microsoft Research and press coverage of the paper.
Source ↗Google
Official disclosure · 2025-08Median Gemini Apps text prompt
- 0.24 Wh
- 0.03 g CO₂e (market-based)
- 0.26 mL of water
Scope: Full serving stack: active AI accelerators, host CPU and memory, idle machines kept for reliability, and data-center overhead. Emissions are market-based, i.e. after Google's clean-energy purchases. Prompt length is not disclosed.
Source ↗Mistral AI
Official disclosure · 2025-07A 400-token response from Le Chat (Mistral Large 2)
- 1.14 g CO₂e
- 45 mL of water
- 0.16 mg Sb-eq of resource depletion
Scope: Life-cycle assessment: includes the manufacturing of servers and other upstream impacts, not only the electricity used to answer. No energy (Wh) figure was published.
Source ↗OpenAI (Sam Altman)
Executive statement · 2025-06Average ChatGPT query
- About 0.34 Wh
- About 0.000085 gallons of water (≈0.32 mL)
Scope: Stated in a blog post by OpenAI's CEO. No methodology, system boundary or model was specified.
Source ↗Epoch AI
Independent research · 2025-02Typical ChatGPT (GPT-4o) query with about 500 output tokens
Scope: Independent first-principles estimate from hardware power draw, utilization and data-center overhead. Corrected the widely repeated older figure of about 3 Wh per query.
Source ↗No per-query figure of their own that we have found: Anthropic, Meta. Know of a publication we have missed? support@co2radar.com
How CO2Radar's estimates compare
Our estimator (methodology v2026.06) computed for the same request as each published energy figure, at PUE 1.2. 3 of 5 published figures fall inside our range. Where they do not, we show by how much rather than hide it.
| Published figure | Published | Output tokens | Our estimate | Result |
|---|
Google Median Gemini text prompt | 0.24 Wh | 500 (assumed) | 0.24–0.48 Wh standard class | Within our range |
OpenAI (Sam Altman) Average ChatGPT query | 0.34 Wh | 500 (assumed) | 0.24–0.48 Wh standard class | Within our range |
Epoch AI Typical GPT-4o query | 0.3 Wh | 500 | 0.24–0.48 Wh standard class | Within our range |
Microsoft Research Typical chatbot query | 0.31 Wh | 300 | 0.14–0.29 Wh standard class | Above our range (by 1.1×) |
Microsoft Research Long reasoning query | 3.91 Wh | 5,000 | 9.00–36.0 Wh frontier class | Below our range (by 2.3×) |
Mistral is not directly comparable. Mistral published 1.14 g CO₂e for a 400-token response as a life-cycle figure that includes hardware manufacturing, and no energy figure. Our operational estimate for the same response on the French grid (55 g CO₂/kWh) is 0.011–0.021 g CO₂e — electricity only. The gap is a difference in what is counted, which is why CO2Radar labels its numbers as operational, location-based estimates.
"Assumed" means the source does not state the response length, so we compare against our default query of 500 output tokens. Full method: methodology.
Frequently asked questions
- Which AI companies have disclosed the energy use of a query?
- As of 2026-11-17, Google is the only major provider to publish a per-prompt energy figure with a documented method: 0.24 Wh for a median Gemini text prompt (August 2025). Mistral AI published a life-cycle assessment giving 1.14 g CO₂e per 400-token response (July 2025) but no energy figure. OpenAI's CEO stated about 0.34 Wh per average ChatGPT query (June 2025) without a methodology.
- Why are the published figures so hard to compare?
- They measure different things. Google reports energy for a median prompt across its whole serving stack and market-based emissions; Mistral reports life-cycle CO₂e including hardware manufacturing; OpenAI's figure has no stated boundary; research estimates assume a fixed number of output tokens. Prompt length, model size, what counts as overhead and how clean energy purchases are counted all change the result.
- How much energy does one AI query use, according to published figures?
- For a typical text query, published figures cluster around 0.2 to 0.35 Wh: 0.24 Wh (Google, median Gemini prompt), about 0.3 Wh (Epoch AI, GPT-4o), 0.31 Wh (Microsoft Research, median) and about 0.34 Wh (OpenAI's CEO). Long reasoning queries use far more — about 3.9 Wh in the Microsoft Research estimate.