Is Claude as bad as a banana?

What we learned trying to measure our AI carbon footprint.

We use Claude every day. Quite a lot of it. And we’re a B Corp that takes sustainability seriously. Those two things don’t sit comfortably together – and we think it’s worth saying that out loud rather than quietly hoping nobody asks.

This post is our attempt to be honest about where we’ve got to, what we don’t know yet, and what we’re doing in the meantime.

First: the banana

There’s a book called How Bad Are Bananas? by Mike Berners-Lee (yes, Tim’s brother) that does something really useful. It gives you the actual carbon footprint of everyday things – your breakfast, your commute, a World Cup – and in doing so makes climate impact feel a bit less abstract and a bit more like something you can actually reason about.

A banana, for instance, has a carbon footprint of 110 grams of CO2. That’s surprisingly low for a tropical fruit that’s been shipped thousands of miles – it turns out bananas grow in natural sunlight, keep well, and travel by boat rather than air freight (which uses roughly 100x the carbon). You can eat a banana without much guilt.

A pint of British milk, by contrast, clocks in at nearly 10 times that figure. Cows require a lot of feed, produce methane, and the land, refrigeration, and processing all add up. A return flight from London to Hong Kong? Around 3.5 tonnes of CO2 – roughly equivalent, we’d later discover, to our entire estimated annual AI carbon footprint.

Which brings us to Claude.

The measurement problem

When we started trying to work out the carbon cost of our AI use, we ran into a wall pretty quickly.

Anthropic and OpenAI currently disclose no meaningful sustainability metrics. None. So if you’re an agency trying to account for your AI emissions honestly, you’re stuck using what’s called a spend-based conversion factor – you take how much you spent on AI tools, apply an average emissions factor from an internationally recognised standard (the Greenhouse Gas Protocol), and get a number.

The problem is that this method was designed before AI existed at scale, and it treats a Claude subscription the same way it treats any other software spend. It doesn’t distinguish between tools with very different energy profiles. It doesn’t account for where the servers are, or how the grid is powered, or whether you’re running a three-word prompt or a 50,000-token context window.

We use it anyway, because it’s what the carbon accountants currently recommend and doing something imperfect is better than doing nothing. But we want to be clear: we know it’s inadequate.

Based on our spend-based estimate, our AI use produces around 3.38 tonnes of CO2 per year. That feels low to us, too.

What we’re doing about it

We’re offsetting more than the estimate

Because we don’t trust the number, we’ve committed to offsetting one tonne of CO2 per month – 12 tonnes over the year, or roughly three times what the spend-based figure suggests. We’re investing that through Ecology into blue carbon removal, specifically mangrove planting, which stores carbon at four times the rate of rainforests. It’s not a perfect solution – offsetting is never a substitute for reduction – but it’s a meaningful precautionary commitment while we work on better measurement.

We’re trying to reduce our usage, not just offset it

Offsetting your way out of excessive consumption isn’t the goal. So we’ve also been working on how we actually use AI day-to-day. A few habits that reduce token usage meaningfully:

• Be specific about how long you want the answer to be
• Start a fresh chat when you switch topics – every message you send references the whole conversation history, so a long session that drifts in subject costs far more than it needs to
• Share IDs directly where possible (like ClickUp task IDs) so Claude doesn’t have to search for them, which burns thousands of tokens
• Use /clear regularly in Claude Code to reset the context window

None of these feel like dramatic sacrifices. But at agency scale, across dozens of people using AI throughout every working day, small habits add up.

We’re building something better

The most exciting part of this is an internal project called Glyph, built by Alfie, one of our apprentices. It’s a dashboard that tracks Claude usage by actual token count, translates that into electricity cost using live UK grid rates, and converts it into grams of carbon weighted by the current grid intensity.

That means instead of “we spent £X on AI, therefore Y tonnes of CO2,” we can see “this session used 29 million tokens, cost roughly £2 in electricity, and produced 9 kilograms of carbon” – the equivalent of about nine pints of milk, to bring things back around.

Glyph is still in development – it currently covers Claude via terminal and VS Code, with web tracking still being worked on – and the methodology is genuinely complex (prompt caching, for instance, changes the carbon calculation in ways that take a full PDF to explain). But it’s already showing us things the spend-based approach never could, including the slightly alarming reality of how quickly context memory accumulates tokens when you’re not paying attention.

Why we’re publishing this before it’s finished

Honestly? We nearly didn’t. It can feel uncomfortable to publish something that’s still a work in progress, especially when the imperfections are right there in the methodology.

But we think transparency while you’re still figuring it out is more useful than a polished story you publish once you’ve got all the answers. If we share this and someone points out that we’ve got something wrong, or that there’s a better approach, that’s more information. That’s more data. We can use it.

So: this is where we’ve got to. We use AI heavily. We know our current measurement is imperfect. We’re offsetting beyond what the estimate suggests, reducing where we can, and building tools to get more accurate over time.

Is Claude as bad as a banana? We genuinely don’t know yet. But we’re trying to work it out.