AI Tool Energy Use: What Vendors Have Published

Vendor publications include a scoped estimate of 0.24 Wh for a median Gemini Apps text prompt but no like-for-like cross-vendor figure, and Chat Picker has not tested energy use.

Sources checked 2 Oct 2026

The vendor pages reviewed for this article do not provide a like-for-like energy comparison of ChatGPT, Claude and Gemini. Only Google supplies a per-prompt environmental figure in those pages: Google Cloud Blog’s environmental-impact page estimates that the median Gemini Apps text prompt uses 0.24 Wh of energy, emits 0.03 grams of carbon dioxide equivalent and consumes 0.26 milliliters of water, as read on October 1, 2026. The reviewed OpenAI accuracy note and Anthropic incorrect-response note give no comparable energy, carbon or water figure, and Chat Picker has not tested the assistants for energy use.

What the vendors document

Google says its environmental-impact methodology includes active model computation and actual chip utilization at production scale. It also accounts for idle chips provisioned for availability, reliability, traffic spikes or failover; host CPUs and RAM; and data-center cooling and power distribution.

The same Google methodology uses Power Usage Effectiveness to measure overhead energy efficiency. Per-prompt emissions use Google’s 2024 average fleetwide grid carbon intensity, while water estimates use its 2024 average fleetwide water usage effectiveness. The point-in-time assessment used May 2025 data. Google says the findings do not represent every Gemini Apps text prompt or future performance, and that an independent third party has not verified the data and claims.

All plan prices below are in U.S. dollars.

ChatGPT: As read on October 1, 2026, OpenAI’s ChatGPT pricing page lists unlimited text chats on Free, with limited uploads, images, voice and deep research. Plus is $20 per month billed monthly. Go and Plus expand tool and upload access, while Pro has three usage tiers; the page does not publish exact message counts. Opt-out from training is available on Free, Go, Plus and Pro.
Claude: The Anthropic pricing page, as read on October 1, 2026, says Free works on the web, desktop and mobile and lists web search, file creation, code execution and memory. Pro is $20 monthly or $17 with annual billing, with $200 up front. Max starts at $100 per month and offers 5x or 20x Pro usage. The page gives no message counts. Opt-out applies to Free, Pro and Max; Team is not trained on by default.
Gemini: Google’s US AI plans page, as read on October 1, 2026, says Free requires a Google Account and includes 15 GB of storage. AI Plus is $4.99 per month with 2x the Free usage limits and 400 GB of storage; AI Pro is $19.99 with 4x the limits and 5 TB. Compute-based limits refresh every five hours up to a weekly limit, and AI credits can extend them. The page does not state a chat-training opt-out setting, so Chat Picker marks it Not verified.

In accuracy notes read on the same date, OpenAI says ChatGPT can be incorrect or misleading and may sound confident when wrong, so important information should be checked against reliable sources. Anthropic says not to rely on Claude as the only source of truth and recommends reviewing cited pages for missing context. These are accuracy cautions, not energy measurements.

This replaces an older page whose statistics and test results lacked sources; those claims have been removed.

What the documentation cannot tell you

Google’s number is an estimate for a median prompt, not a reading from your account. It cannot establish the joules per token, accelerator model, idle-chip allocation, regional grid mix, electricity cost or future footprint for your workflow. It also cannot connect lower energy use with better answers.

Your own trial can test output quality and plan behavior, but it is not an energy test unless you add separate metering. Record missing energy fields as unknown rather than estimating them. Chat Picker’s method page says it relies on vendor-published facts, marks unconfirmed figures Not verified, and has no quality, speed, accuracy, reliability or benchmark results of its own.

How to check it yourself

Use identical inputs and separate chats. Keep the plan, device and account settings as consistent as possible, and do not infer energy use from answer length alone.

Separate training from inference. Give the assistant: “Does a per-prompt energy estimate measure model training, one inference request, or both? State its scope, list the included infrastructure, and mark every missing input.” Look for a clear distinction between upfront training and serving. Record each figure, date and stated scope, then confirm it against vendor documentation.
Test a repeatable task. Give the assistant: “Classify each task by owner and status, marking missing fields: Copy editor — draft website copy — status not stated; Accessibility reviewer — review accessibility — status complete; Support editor — update the help center — status blocked.” Look for correct extraction and unsupported assumptions. Record omissions, added claims and whether each field can be verified.
Check repeatability and visible limits. Repeat the same classification request in separate new chats and, if useful, on each plan you actually use. Look for output changes, cutoff notices, rate-limit messages and any displayed usage counter. Record each answer and the interface text; write not shown when no counter appears. These observations do not establish joules per token.
Separate output from infrastructure cost. Give the assistant: “Before estimating the GPU cost of an AI workload, list the billing and hardware facts required. Separate what a chat interface can reveal from what only a dashboard or provider can reveal.” Look for accelerator type, utilization, input and output volume, region and billing rates. Record which items are visible to you and which require provider data.
Check carbon and water context. Give the assistant: “Before converting an AI service’s electricity use into emissions or water use, list the location, grid and time-period inputs required. Explain why fleet averages may not match a particular workload.” Look for a distinction between energy, carbon dioxide equivalent and water consumption, and between averages and local conditions. Record the source and date for every factor; leave local values unknown if they are not published.

Which rows of the comparison matter

Open the three-way comparison matrix and choose the rows that define your trial conditions: Free plan, Main paid plan, Heavy-use plans, Context window in the app, How usage limits are described and Training on your chats. Add Low-cost tier or Team plan if either matches your use.

Those rows help keep plan conditions consistent. They do not convert price, context size or usage allowance into energy. Treat a blank cell marked Not verified as unresolved rather than filling it by inference.