2026 AI Environmental Impact: What Compute and Carbon Figures Exist
Explains the available per-prompt estimate of 0.24 Wh, 0.03 gCO2e and 0.26 mL and provides an afternoon audit plan for AI environmental claims.
Sources checked 3 Oct 2026
Among the vendor pages reviewed, only Google’s Measuring the Environmental Impact of AI Inference gives a per-prompt estimate: 0.24 Wh of energy, 0.03 gCO2e, and 0.26 mL of water for the median Gemini Apps text prompt, as read on October 1, 2026. The reviewed OpenAI and Anthropic pages, including OpenAI’s truthfulness guidance and Claude’s incorrect-response guidance, give no comparable per-prompt environmental figure. Chat Picker has not tested the assistants’ compute use, energy use, water use, or emissions.
Know these limits before you start
Google describes its estimate as a point-in-time analysis using May 2025 data, not a value for every prompt or future performance. Its boundary covers active computation and achieved chip utilization, idle provisioned chips, host CPUs and RAM, accelerators, cooling, and power distribution. Google uses Power Usage Effectiveness for overhead energy efficiency, 2024 average fleetwide grid carbon intensity for emissions, and 2024 average fleetwide water usage effectiveness for water estimates. The page says the data has not been independently verified. The reviewed pages do not provide comparable vendor figures for training carbon, delivered TOPS/W, lifetime footprint, or offset and renewable-energy accounting.
In the same Google post, read on October 1, 2026, Google reports that over a recent 12-month period the median prompt’s energy footprint dropped 33x and its total carbon footprint dropped 44x.
Treat an assistant’s environmental explanation as a claim, not a measurement. OpenAI says ChatGPT can sound confident while being wrong and recommends checking important information. Anthropic says Claude can produce convincing but ungrounded statements and should not be the only source of truth. This rewrite removes the older page’s unsourced statistics and test results.
The test plan
Run this as a documentation audit, not a physical measurement. Send each request unchanged to all three assistants and record the plan, visible model label, date, and tool state.
-
Prepare one plain-text request. Send: “State the environmental impact of one text prompt. Separate measurements from estimates, and give the boundary, location, period, energy source, offset treatment, source, and verification status. Write ‘not stated’ for anything absent.” Look for a unit and traceable citation, not a broad efficiency claim. Log the exact reply and links.
-
Prepare the first reply and its cited pages. Send: “For this exact prompt, separate effects attributed to model architecture from active chip utilization, idle capacity, host CPUs and RAM, cooling, and power distribution. Identify any PUE, grid-carbon, and water method, with its date and source.” Look for each boundary item and attribution. Log present, absent, and conflicting statements separately.
-
Prepare a hardware request. Send: “For the exact deployed system, state whether chip and TOPS/W are disclosed. Separate nameplate TOPS/W from achieved production utilization, cite the source and date, and write ‘not stated’ if the service does not expose them.” Look for a supported system name, not an inferred one. Log whether hardware specifications and workload efficiency stay separate.
-
Prepare a lifecycle worksheet with rows for training, one inference, repeated use, and a full lifetime. Send: “For each row, give any documented figure, allocation method, period, boundary, and source. Do not turn a total into a per-prompt value without a stated allocation.” Look for training-carbon debt and lifetime totals. Log every conversion and leave missing values blank.
-
Prepare an energy-and-claims worksheet. Send: “List evidence for renewable electricity, data-center location, PUE, water consumption, carbon offsets, and reporting period. Distinguish consumed energy or emissions from avoided or compensated emissions, and cite each source and date.” Look for location and time, not only a carbon-neutral label. Log instruments separately from consumption.
-
Prepare each answer with its cited source pages. Send: “Check every cited page and quote the passage supporting each number. If a number, scope, or date is absent, say so.” Open the links yourself as well. Look for exact support and third-party verification. Log matches, mismatches, inaccessible pages, and unverified claims; never turn an inaccessible source into zero.
How to read your results
Read across rows, not down a single favorite. A number is traceable only when your log captures its unit, service and model scope, geography, period, method, whether it is measured or estimated, and source date. Keep training totals, per-prompt serving estimates, and lifetime allocations separate; adding or ranking them without a shared boundary would create a comparison the evidence does not contain.
Treat “not stated” as unknown, not zero. If answers conflict, keep both, open the sources, and mark the unresolved item. Choose a plan from the documentation you can trace, the context and usage room you need, and the data setting that fits your work—not from a claim that one assistant has lower impact.
Where the plans differ
The Claude vs ChatGPT vs Gemini matrix is where to check the documented plan rows. As read on October 1, 2026, the low-cost row lists ChatGPT Go at A$13/month on OpenAI’s Australian pricing page, no Claude tier between Free and Pro on Anthropic’s pricing page, and Google AI Plus at $4.99/month on Google’s US plans page; a US-dollar Go price was not verified. The main paid row lists ChatGPT Plus at $20/month, Claude Pro at $20/month or $17/month with $200 up front, and Google AI Pro at $19.99/month. Recheck the vendor page before subscribing.
For the context row, OpenAI lists 27K, 54K, and 128K for Instant models by plan, plus 256K and 400K for reasoning models; Anthropic lists up to 1M on every Claude plan, varying by model. No Google app token figure was verified, though its plans page lists uploads of up to 1,500 pages. All three pages were read October 1, 2026. Their pricing pages do not publish exact message counts for ChatGPT or Claude, while Google’s compute-based limits refresh every five hours up to a weekly limit.
Before using a test account, check the training-on-chats row. OpenAI says opt-outs are available on Free, Go, Plus, and Pro; Anthropic says they are available on Free, Pro, and Max, while Team is not trained on by default. The Google Gemini setting was not verified. These settings were read October 1, 2026.