One month of heavy AI use

What heavy means

Measured, not guessed

One person's use of Claude Code, an AI coding agent, was logged for two weeks, 14 to 27 September 2026, and scaled to a month. They were on the Claude Max 5x plan, at or near its weekly limit, so this is close to the most that plan allows.

• 35,000requests to the model in a month
• 320,000tokens of conversation behind every token it wrote, on average (a token is a word or part of one)
• 11 billiontokens re-read from a cache in a month

An agent re-reads the whole conversation every time it acts, and these conversations run long. A chat prompt is short. That is what makes heavy use heavy.

Why the range is so wide

Anthropic publishes nothing

No energy, carbon or water figure for Claude, per request or in total, as of 28 Sep 2026. So every figure here is estimated from other people's measurements. Three unknowns decide it, and one thing is left out.

So a month is 8.2 kg to 670 kg, answering only: the high end is 82 times the low. Where the evidence can't decide, the central estimate takes the less flattering side.

All at once

Both are true

A daily chat habit is a speck: about 0.11 miles of driving a month. A heavy month of agent work is about 1,100. A fair argument about AI needs both.

Petrol-car miles a month, by carbon. Each amber square's area is to scale against the largest; a speck is drawn one pixel wide so it can be seen at all. Every number and every source is below.

Where the needle moves

Not in one person's subscription

• 16%of the world's consumption emissions in 2019 came from the richest one per cent (77 million people): as much as the poorest two-thirds (5 billion people).
• 80%of the world's fossil CO₂ since 2016 is linked to 57 fossil-fuel and cement producers, most of it from customers burning what they sell.
• 23%rise in Microsoft's reported emissions since 2020, which it puts down to growth "such as AI and cloud expansion".
• 2004when BP launched its "carbon footprint calculator", part of a campaign to make climate change about individuals' habits.

What actually helps

Aim at the big levers

• Together. 60% of the cuts on the UK's official path to 2040 come from electrifying things and cleaning up the power supply: planning, grids and policy. England only lifted its de facto ban on onshore wind in 2024. Vote and campaign on it.
• At home. Households' own choices make 33% of those cuts, helped or blocked by policy. The biggest: living car-free, 2.4 tonnes a year; a plant-based diet, 800 kg a year; one fewer return flight across the Atlantic, 1.6 tonnes. Recycling saves 210 kg a year. A heat pump cuts heating emissions by up to 70%.
• Your money. Pensions and savings are invested for you, and you can choose how: even Nest, the government-backed workplace pension, offers funds for "different beliefs".
• Heavy AI use. On a plan used to its weekly limit, the footprint is roughly set by the limit, so working well buys more work, not more energy. The rest is the provider's: its efficiency, the power it buys, and publishing its figures. A smaller plan cuts it, like driving less; whether the work is worth it is each user's call.

This page's view: a heavy user's footprint is real, but it is on the scale of running a car, not of the levers above. Until those change, blaming the people who subscribe won't move the needle.