AI Tool Environmental Impact: Published Energy and Water Figures

The reviewed vendor pages provide a scoped per-prompt estimate for the median Gemini Apps text prompt but no comparable figures for ChatGPT or Claude.

Sources checked 2 Oct 2026

Only Google’s environmental-impact analysis of AI inference, as read on October 1, 2026, provides a published per-prompt environmental figure among the pages reviewed: it estimates that the median Gemini Apps text prompt used 0.24 watt-hours of energy, emitted 0.03 grams of carbon dioxide equivalent and consumed 0.26 milliliters of water in a point-in-time assessment using May 2025 data. The specific OpenAI accuracy note and Anthropic incorrect-response note reviewed for this page give no comparable per-prompt figure, and Chat Picker has not tested or measured the assistants for environmental impact.

What the vendors document

Google says its method covers active model computation and actual chip utilization at production scale, plus idle chips provisioned for availability, reliability, traffic spikes or failover. It also includes host CPUs and RAM, cooling and power distribution. Power Usage Effectiveness, or PUE, measures overhead energy efficiency. The figure is therefore an operational estimate, not the electricity use of one accelerator.

The same Google analysis, as read on October 1, 2026, gives 0.10 Wh, 0.02 gCO2e and 0.12 mL per median prompt when counting only active TPU and GPU consumption; Google characterizes that as substantially underestimating the real operational footprint. In that source, as read on the same date, the median prompt’s energy footprint dropped 33x and its total carbon footprint dropped 44x over a recent 12-month period.

The analysis uses May 2025 data in a point-in-time assessment. Per-prompt emissions combine energy per prompt with Google’s 2024 average fleetwide grid carbon intensity; the water estimate uses Google’s 2024 average fleetwide water usage effectiveness. Google says the results do not represent all Gemini Apps text-generation prompts or future performance, and that energy per median prompt can change with models, architecture and user behavior. It also says the data and claims have not been verified by an independent third party.

As read on October 1, 2026, OpenAI’s pricing page says Free has limited messages with uploads but no exact count; Anthropic’s pricing page says Pro offers more usage than Free, also without a message count. OpenAI says an opt-out from training on chats is available on Free, Go, Plus and Pro; Anthropic says it is available on Free, Pro and Max, while Team is not trained on by default. These are data controls, not footprint measurements.

OpenAI’s accuracy note says ChatGPT may sound confident when wrong and important information should be verified. Anthropic’s note says Claude should not be the only source of truth and original sources should be reviewed. Gemini Apps’ source help says the Sources button appears when sources are available and may show public websites, uploads or connected Workspace documents and email. Check generated environmental summaries against the originals.

What the documentation cannot tell you

The Google result is scoped to a median text prompt in a defined assessment. It cannot tell you the footprint of your full conversation, every prompt, a particular subscription or a future system. The reviewed pages also provide no cross-vendor training-energy total, comparable inference benchmark, offset or renewable-energy claim, hardware-waste total or 2025 per-tool environmental rating.

Your trial can reveal prompt length, tool use, citation behavior and answer consistency. It cannot calculate actual electricity or water use without metered infrastructure data and a defined lifecycle boundary. Do not infer footprint from response speed, length, quality, plan price, context window or training opt-out.

An earlier version of this page contained statistics and test results without sources; those items have been removed.

How to check it yourself

Use the same wording for each assistant and record whether browsing or source tools were available.

Recover the exact claim. Give each assistant: “Open Google’s page titled ‘Measuring the environmental impact of AI inference.’ Report the median Gemini Apps text-prompt estimate for energy, carbon dioxide equivalent and water. Include the data month, operational boundary and exact source link; say if an item is absent.” Look for “median,” “estimated,” app scope and date. Write down each value with its unit, scope and source read date.
Separate accelerator use from operations. Give each assistant: “Using Google’s page titled ‘Measuring the environmental impact of AI inference,’ explain the difference between counting active TPU and GPU consumption and using the broader operational boundary. List the included components and do not convert units.” Look for idle chips, host CPUs, RAM, cooling and power distribution. Write down the included and excluded boundaries and whether the source calls the narrower count an underestimate.
Check the reported trend. Give each assistant: “On Google’s page titled ‘Measuring the environmental impact of AI inference,’ identify any reported change over a recent 12-month period. State the exact metric, multiplier and caveats about future performance or changing models, architecture and behavior.” Look for whether energy and carbon are treated separately from water. Record each metric and caveat without extrapolating it.
Look for a comparable vendor figure. Give each assistant: “Search only official OpenAI, Anthropic and Google pages for a per-prompt energy or water figure for ChatGPT, Claude or Gemini. Give the exact metric, product scope and source, or state that the official page checked does not provide it. Do not estimate.” Look for app versus API scope, median versus average, and estimate versus measurement. Record tool access and the exact citation.
Audit broader environmental claims. Give each assistant: “Find official evidence for training energy, lifecycle carbon, data-center water, hardware waste, renewable-energy matching, offsets, a 2025 per-tool environmental rating, or a regulatory or future claim about ChatGPT, Claude or Gemini. Quote only direct evidence and mark missing evidence as not stated.” Look for the accounting boundary, date, method and any third-party verification. Write down each claim, its evidence and remaining gaps.

Which rows of the comparison matter

Use the full ChatGPT, Claude and Gemini comparison matrix to hold workflow context constant, not to create an environmental score. The relevant rows are:

Plan access: Free plan, Low-cost tier, Main paid plan, Heavy-use plans and Team plan.
Workload boundaries: Context window in the app and How usage limits are described.
Data handling: Training on your chats.

Each matrix figure carries its vendor source and read date. An unconfirmed cell is marked Not verified and left blank. These rows show what access and controls apply, but they do not establish training energy, per-prompt water use or a lifecycle footprint. Keep missing values as gaps rather than turning price, context size, usage wording or a training control into an environmental estimate.