How much energy and CO₂ does an AI agent use per task or session?

Far more than a chat prompt, because a task is many model calls and each one re-sends the whole working context. SOMA AI estimates that an illustrative agent task of 10 million billed tokens, with 91% of input read from cache, uses about 135 Wh on a mid-size model and 504 Wh on a frontier model. At the world-average grid that is about 98 g and 382 g CO₂e. One chat exchange of 1,210 tokens on the same mid-size class comes to about 0.099 Wh, so the task equals roughly 1,400 chat exchanges. A widely reported public estimate, by the climate scientist Zeke Hausfather, puts each prompt he typed into Claude Code at around 150 Wh, about 600 times a median chat prompt. Both are estimates. No provider we checked publishes energy per agent task.

A chat is one request and one answer. An agent loops. It plans, reads files, runs a command, reads the output and calls the model again, and every call carries the conversation so far. Anthropic's documentation says Claude Code "sends your full conversation with every request", and that each round of tool use sends another request carrying the tool results (Anthropic, Claude Code documentation, "Manage costs effectively", read 11 November 2026). Four things add up.

• Many calls per task. At SAEF's default of 85,000 tokens per agent request, 1 million tokens is about 12 model calls and 100 million is about 1,176.
• Context re-read on every call. Files, instructions and earlier turns are sent again each time, so input dwarfs output. In the dataset default, output is 0.5% of the tokens of an agent request.
• Tool output. Test logs, search results and file contents enter the context and are then re-read on every later call.
• Subagents and teams. Each subagent or teammate has its own context. Anthropic states that agent teams use about 7 times more tokens than a standard session when teammates run in plan mode (Anthropic, Claude Code documentation, "Manage costs effectively", read 11 November 2026).

Most of the re-read context is served from the provider's prompt cache, which costs far less compute than processing it fresh. That is why cached input has to be counted separately. SAEF weights a cached input token at 0.05 and an uncached one at 0.3 of an output token. With no cache credit, the same 10 million-token mid-size task would estimate 534 Wh, about 3.9 times more.

• One user's Claude Code use, estimated. Hausfather reports that over 8 weeks his Claude Code used 3.2 billion tokens across 1,138 typed prompts, and estimates about 170 kWh of data-centre electricity, with a range of 70 to 330 kWh across methods. That is around 150 Wh per prompt, roughly 600 times (250 to 1,200) a median chat prompt, benchmarked on Google's median Gemini text prompt and OpenAI's average ChatGPT query. His median session comes to about 0.6 kWh and his average day to 3.0 kWh. About 96% of the tokens were cache reads (Zeke Hausfather, The Climate Brink, 5 August 2026). He notes that nobody outside the labs knows the true energy per token of a frontier model.
• A second Claude Code estimate. Simon P. Couch, who builds data-science software at Posit, writes on his personal blog that, working from Epoch AI's per-query energy analysis, he estimates his median Claude Code session at 41 Wh and his median day at 1,300 Wh. He calls it napkin math (Simon P. Couch, "Electricity use of AI coding agents", 20 January 2026). Hausfather uses Couch's per-token figures as one of his three methods, so the two estimates are not independent.
• Tokens per coding task. A study of eight frontier models running the OpenHands agent on SWE-bench Verified finds that agentic coding uses about 1,000 times more tokens than code reasoning or code chat, that input tokens drive the cost, that the same task can vary up to 30 times between runs, and that Kimi-K2 and Claude Sonnet 4.5 use on average over 1.5 million more tokens per task than GPT-5 (Bai et al., "How Do AI Agents Spend Your Money?", arXiv 2604.22750, v3 2 October 2026).
• Provider statements. Anthropic publishes cost, not energy. Its documentation gives an average Claude Code cost of about $13 per developer per active day across enterprise deployments, below $30 per active day for 90% of users, and says one debugging session can consume more than a day of chat (Anthropic, Claude Code documentation, "Manage costs effectively", read 11 November 2026). In what we checked on 11 November 2026, none of Anthropic, OpenAI, GitHub, Cursor or Cognition published an energy or emissions figure per agent task.

As a cross-check, SOMA ran Hausfather's token count through SAEF. 3.2 billion tokens with 96% cache reads give about 133 kWh on a frontier (class C) model, inside his 70 to 330 kWh range. On a mid-size (class B) model, the class SAEF assigns to Claude Sonnet, they give about 36 kWh, below his range. The post names no model, so the class is SOMA's assumption and it decides whether the two agree. The class C figure is about 2.4 kWh a day for around 57 million tokens a day. The two methods use different per-token assumptions, so at most this shows the same order of magnitude.

The table applies SAEF to three task sizes. The assumptions are the dataset's agentic defaults. Output is 0.5% of tokens, 91% of input is read from cache (the share in the one sample session shown in Anthropic's cost documentation, not a survey), the GPU is an H100 and the boundary covers electricity, hardware, the building and amortised training. Energy does not depend on the grid. CO₂e is shown on three grids. These are the cleanest grid that hosts a cloud region in the dataset, Quebec (Canada), the world average and the most carbon-intensive, Bahrain, out of the 46 grids that host an AWS, Azure or Google Cloud region.

Model class moves energy about 3.7 times between mid-size and frontier. The grid moves CO₂e about 4.3 times between Quebec (Canada) and Bahrain on the mid-size class, less than the gap in grid intensity because hardware, the building and training do not depend on the grid. The chat comparison uses one exchange of the same class. The token sizes are illustrative, not a statistic of how long tasks run. Use your own token counts.

• Count billed tokens, not seats. Two developers on the same seat can differ by many times in tokens. Where agents run on an API account or a cloud provider, export tokens per model from the usage reports, the Claude Code Analytics API or OpenTelemetry, as Anthropic's documentation describes (Anthropic, Claude Code documentation, "Manage costs effectively", read 11 November 2026).
• Keep cached input as its own column. Most agent input is cache reads. If the export splits them, enter them separately. If not, the estimate takes no cache credit and runs high.
• Record the model and the region. They move the estimate more than anything else on this page. Where a tool shows only seats or request counts, use the per-request figures on the coding assistants page and say so.
• Use the levers. Clearing context between tasks, smaller models for subagents and keeping the cache warm cut tokens and effective tokens. The page on cutting the AI footprint sizes each lever.
• Report it in Scope 3, Category 1. Agent use bought from a provider is a purchased service under the GHG Protocol. The method is in the Scope 3 guide, and the free calculator takes tokens with cached input.

• That any figure is a measurement. SOMA estimates from the SAEF factors, and Hausfather and Couch also estimate, from token counts.
• That any named tool uses a given model class, region or number of tokens per task. The task sizes are illustrative.
• That one tool is cleaner than another. SAEF has no term for the tool brand, only tokens, model class and region.
• That the 600 times figure applies to every user. It is one person's usage, per typed prompt, against a median chat prompt.
• That agents are worse for the climate than the work they replace. This page does not estimate the alternative.
• That agents are a large share of a company's footprint. That depends on how many people run them and how much, and this page does not estimate it.
• Market-based figures. All CO₂e here is location-based.

The SAEF method was accepted for publication in Sustainable Production and Consumption. The equations are on the methodology page.

• What is the carbon footprint of AI coding assistants like Claude Code, GitHub Copilot and Cursor? Per completion and per agent request
• How much more energy do reasoning models use than standard chat? Why hidden reasoning tokens matter
• What is the carbon footprint of Claude, and how does it compare with ChatGPT? Haiku, Sonnet and Opus per exchange
• Where does an AI query emit the least? The carbon footprint of AI by country. The same exchange on 210 grids