AI agent energy has now been measured over eight weeks of real Claude Code usage, and a single request sent to an agent costs 150 Wh. That is roughly 600 times what a standard chat message costs, which vendors put at 0.24 to 0.34 Wh. The gap comes from the number of model calls each typed instruction fires off in the background.
Key Takeaways
- 1,138 typed requests over eight weeks produced more than 14,000 model calls and 3.2 billion tokens.
- Measured consumption lands at roughly 170 kWh over the period, within a 70 to 330 kWh range.
- 96% of the tokens counted are cache reads, not freshly generated content.
Have an AI Sum Up This Article
ChatGPTOne hundred and seventy kilowatt hours for eight weeks of code
The protocol is simple, and that is what makes it useful. To pin down AI agent energy, a climate scientist instrumented his own use of Claude Code for eight weeks, then converted billed tokens into energy.
The raw count gives 1,138 hand typed requests. Those instructions fired more than 14,000 model calls and moved 3.2 billion tokens.
The estimated energy bill settles around 170 kWh over the period, with an acknowledged uncertainty between 70 and 330 kWh. Broken down daily, that is 3 kWh a day on average, in a 1.2 to 5.9 range.
The order of magnitude lands once you translate it into household terms. That usage draws more than a pair of refrigerators running non stop, for one developer.
Extrapolated across twelve months, the trajectory gives roughly 1.1 MWh and close to 370 kg of CO2 equivalent. The full analysis and its calculation method are published on The Climate Brink.
The finding lines up with what teams already see on the billing side. We saw it when Microsoft dropped Claude Code after burning its entire AI budget, without anyone at the time tying the cost overrun to an energy overrun.
Cache reads account for most of the bill
The mechanic comes down to one figure. Of the 3.2 billion tokens counted, 96% are cache reads rather than content generated for the first time.
An agent rereads its own context at every step. It pulls back the history, the open files, the previous tool output, then adds its increment. Multiply that by the number of steps in a task and the volume swells without the user typing another word.
That loop is what opens the gap with chat. A question asked in a conversation window is worth one call. An instruction handed to an agent can be worth fifty, most of them invisible.
The ratio of twelve model calls per typed request lays bare the cost structure. The user perceives one action, the system executes a dozen, and both billing and consumption track the second number rather than the first.
The cache paradox deserves a mention in any AI agent energy calculation. Caching was introduced to cut the cost of rereading context, and it did exactly that at the level of a single call. What it mostly did was make a multiplication of calls economically acceptable.
The measurement also explains why speed gaps between agents cost so much. We noted it on usage benchmarks when the fastest coding agent turned out to be the priciest one, that speed being bought through parallel calls.
Every capability added to an agent stretches the chain further. Direct web access, as when Claude Code shipped its own built-in browser, multiplies context round trips with every page it opens.
More articles on Horizon
- Amazon Data Center in Texas Will Set a CO2 Record
- Kimi K3 Escape Sent It Straight to GitHub
- Claude Fable 5 Cuts Biology Blocks by 85%
Why the numbers labs publish no longer hold
The figures vendors communicate cover the chat request. Google puts 0.24 Wh per query on Gemini, OpenAI cites 0.34 Wh for ChatGPT.
Those values are not wrong, they measure something else. They describe the cost of one isolated call, at a point where the agentic share of traffic is exploding and the relevant unit is no longer the call but the task.
For product teams, the consequence is immediate. A carbon report that estimates AI agent energy from the chat number understates a real deployment by three orders of magnitude, and capacity projections built on that base break in the first quarter.
On the competitive side, pressure shifts toward context efficiency. Whichever lab cuts cache rereads without degrading task quality wins on price, on latency and on footprint at once, three commercial arguments that used to pull in different directions.
The method used here is worth borrowing internally. Instrumenting eight weeks of real usage costs almost nothing and yields a clean calculation base, where most teams still reason on order-of-magnitude figures handed to them by their vendor.
The acknowledged 70 to 330 kWh range is the weak point of the exercise, and the author says so plainly. It stems from the absence of public data on energy consumed per token served, a figure no lab publishes today.
The finding also tempers some adoption talk. Mark Zuckerberg recently conceded that agents were running slower than expected at Meta, and the slowness observed in production now has a numeric counterpart on the energy side.
Our read is that the next reporting line worth watching is not the count of billed tokens. It is the ratio between tokens reread and tokens produced, the only indicator that tells you whether an agent is working or circling inside its own context.
Follow the story on Horizon.
