Chat Picker

2026 AI Tool Energy Efficiency: Reading the Published Numbers

Google estimates 0.24 watt-hours of energy for the median Gemini Apps text prompt, while the OpenAI and Anthropic pages reviewed in October 2026 publish no comparable per-prompt figure.

Sources checked 3 Oct 2026

For the median Gemini Apps text prompt, Google estimates energy use of 0.24 watt-hours, carbon dioxide equivalent emissions of 0.03 grams, and water consumption of 0.26 milliliters; its point-in-time assessment used May 2025 data (Google Cloud’s environmental-impact article, as read on October 1, 2026). The OpenAI pricing page and Anthropic pricing page, also read on October 1, 2026, publish no comparable per-prompt figure, so the published record does not support an energy-efficiency ranking. Chat Picker has not tested or measured the assistants (sourcing method); this test plan helps you audit the numbers and collect observable results in one afternoon.

Know these limits before you start

Google says its methodology covers energy and water used in active model computation and actual chip utilization at production scale, plus energy from provisioned idle chips, host CPUs and RAM, accelerators such as TPUs and GPUs, cooling, and power distribution. Power Usage Effectiveness is its metric for overhead energy efficiency. The calculation uses energy per prompt and Google’s 2024 average fleetwide grid carbon intensity for emissions, plus its 2024 average fleetwide water usage effectiveness for water consumption. Google says the data and claims have not been independently verified, the finding does not represent all Gemini Apps text prompts or future performance, and the values can change as models, architecture, and use change. Because its geographic factors are fleetwide, this is not a region-specific result.

The plan descriptions, as read on October 1, 2026, say OpenAI does not publish exact message counts, Claude describes more usage without a message count, and Google’s compute-based limits refresh every 5 hours, up to a weekly limit. Check the context and upload ceiling for your chosen plan before preparing each run.

For chat data, OpenAI offers an opt-out on Free, Go, Plus, and Pro, while Claude offers one on Free, Pro, and Max and says Team is not trained on by default. Google’s plans page does not state a Gemini chat-training setting. OpenAI and Claude also warn that responses can be incorrect or misleading and should be checked against reliable sources (OpenAI guidance; Claude guidance). Keep non-sensitive material out of this test.

The test plan

Use fresh sessions, the same device and network, and non-sensitive files. Create one log row per prompt with columns for the plan, prompt ID, start and end time, response location, tool use, interruption, task checks, environmental entry, and allocated cost.

  1. Set the energy column. Prepare a plain-text document containing a title, date, deadline, and five numbered steps. Paste it into each assistant and give this instruction: “Summarize the document in five bullets, then quote the sentence supporting each summary point.” If an energy, carbon, or water readout appears, record its source, scope, and read date. Copy Google’s figures only into the stated median Gemini text-prompt case, and mark the other vendors “not published,” never zero.

  2. Time repeated runs. Use the same document and instruction during your expected peak period and again off-peak, with three fresh runs in each period. Time from submission to the completed response with the device clock. Look for queue notices, tool activity, retries, or truncation. Record every duration and interruption. Label these as session observations, not energy measurements or vendor benchmarks.

  3. Check task quality. Make an answer key from the document, then append: “What is the title, what date does it state, and which numbered step contains the deadline? Quote the supporting sentence for each answer and mark anything not supported.” Look for correct extraction, supported conclusions, and unsupported claims. Record pass or fail for each question and quote any error. Do not turn the log into an overall accuracy score.

  4. Keep carbon geography separate. Record your country or region as a workload label. With Google’s methodology pasted above the prompt, add: “Do not estimate local emissions. State whether this source gives a region-specific figure and identify its scope.” Check the response against Google’s published method. Record the fleetwide scope and mark a local value for each assistant as not published. Do not convert the published Wh value into local carbon without measured energy and a sourced local grid factor.

  5. Allocate subscription cost. Run the same five-prompt battery on every plan that can accept the input and complete the battery within its limits. For a paid plan, divide its monthly price by the number of completed prompts and label the result “allocated test cost per completed prompt,” not API price or marginal billing. Record the billing cadence, completed prompts, failures, and usage notices. Record a free plan as free; that does not establish lower energy use.

  6. Audit hardware scope. Paste Google’s article and ask: “List the infrastructure included in the methodology, then say whether it provides a prompt-level GPU, TPU, or custom-ASIC split. Do not infer a missing split.” Verify the response against the article. Record active and idle compute, host CPU and RAM, cooling, power distribution, and Power Usage Effectiveness as methodology categories. Mark a prompt-level accelerator split or cross-vendor hardware comparison as not published. A hardware category does not prove which chip served your prompt.

How to read your results

Start with requirements rather than an average. A plan does not fit your check if the input exceeds its documented context or upload limit, the battery reaches a usage limit, or its data setting does not meet your needs. Among plans that pass those checks, review your own task results and latency records; treat allocated cost as a planning aid.

Keep the energy field categorical: Google’s scoped estimate or not published. Missing data is not zero, and one vendor’s estimate cannot become a three-vendor comparison. If no option meets your requirements, report that result rather than inventing a winner.

Where the plans differ

Use the Claude vs ChatGPT vs Gemini matrix to check the rows that affect this test: Free plan, Main paid plan, Heavy-use plans, Team plan, Context window in the app, How usage limits are described, Training on your chats, and Ads. Heavy-use, team, and limit rows matter when the battery is intensive or shared.

For an individual paid plan, ChatGPT Plus is $20 per month, billed monthly; Claude Pro is $20 monthly or $17 monthly with annual billing and $200 up front; and Google AI Pro is $19.99 per month, all read on October 1, 2026. These are subscription prices, not per-prompt prices. For the Ads row, OpenAI says Go may include ads; the reviewed Claude and Google plan pages make no ads statement. Blank matrix cells mean “Not verified,” not zero or unavailable; the sourcing method explains the rule.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants