Cheaper Is Greener: Every Token You Cut Is Energy You Don't Burn

The same lever that lowers your AI bill lowers your environmental footprint. Not two goals. One. Every efficiency trick I write about here (cache the thing, right-size the model, stop re-sending the same bytes) is simultaneously a cost story and a sustainability story, because they're the same story told twice.

That's the thesis this whole blog rests on. Including the part where the popular version of it is wrong.

What actually happens when a model emits a token

A token isn't free and it isn't abstract. When a model writes one, a GPU somewhere runs a forward pass. That draws power. The chips get hot, so the datacenter spends energy (and often water) to cool them. Every token you generate is a small physical event in a building you'll never see.

So the chain is: tokens → compute → energy → carbon and water. The mechanism is not in dispute. More tokens means more compute means more kilowatt-hours means more grams of CO₂ and more liters evaporated. That part is just physics and plumbing.

The magnitudes are where it gets interesting — and where most of the internet is quietly wrong in both directions.

The number you shouldn't trust

Here's the doom version you've seen: every ChatGPT query drinks a bottle of water and boils a kettle. Here's the dismissive version you've also seen: it's a rounding error, ignore it. They can't both be right, and honestly, neither is.

In August 2025, Google published metered production numbers for Gemini: real measurements from across their serving fleet. The median text prompt drew 0.24 watt-hours, emitted 0.03 grams of CO₂e, and evaporated 0.26 milliliters of water. About five drops. Energy-wise, that's microwaving something for one second. (First-party numbers from the operator's own fleet, about as good as this data gets. Two caveats: it's Google's median prompt on Google's hardware, not a universal constant, and the water figure counts on-site cooling only, not the water used to generate the electricity.)

Now the other end. A 2025 benchmark that ran 30 models found energy per query spanning 0.45 Wh to 33.6 Wh (a spread of more than 70×), with carbon from 0.3 to 14 grams. (External research; the high end is a heavy reasoning model chewing a long prompt, so treat it as the ceiling, not the norm.) Water estimates run even wider: from Google's 0.26 mL per prompt up to about 500 mL for a conversation of 10–50 responses in an earlier UC Riverside study, a spread of more than a thousandfold. (Different models, different datacenters, different units of work, different accounting. That's exactly the point.)

When your credible estimates disagree by 1000×, anyone quoting you a single confident number is selling something. The honest read is the boring one: a single query is genuinely small. Hannah Ritchie's sober math puts even 100 queries a day at about 0.1% of the average American's electricity, and on Google's new metered numbers she reckons the older scary energy estimates were more than 10× too high. (That 10× is about energy — a different axis from the 1000× disagreement over water accounting above.)

So if you came here worried that asking Claude a question is an environmental sin, relax. That's not where this goes.

Where it actually matters: volume × model choice

Here's where it goes. The per-query number being small is a trap, because it makes you look at the wrong unit.

Two multipliers turn "small" into "matters."

The first multiplier is model choice. That spread is a lever. Run the same task on a small, right-sized model instead of a giant reasoning model and you move a long way down that range (though part of the spread is prompt length, not model alone). The sharper version comes from a peer-reviewed FAccT paper by Luccioni and colleagues: general-purpose generative models cost well over an order of magnitude more than task-specific ones doing the equivalent job — 20–30× in their measurements. Picking the smaller model that does the job is the biggest footprint lever most people actually control, and it's free. The dollars you save routing work to a smaller model are carbon you save in the very same motion. The cost lever and the carbon lever are the same lever.

The second multiplier is volume, and it's the one the consumer discourse misses. Everyone argues about a human typing one prompt. But that's not where the tokens are. The tokens are in automation: agents, pipelines, batch jobs, things that run without a person watching.

Let me give you a real number from our own setup. One autonomous run of our planning skill (a single /plan on a single ticket) cost 2.40 million effective-billable tokens on its main thread. One run. A person would have to sit and type queries for a very long time to reach that. Agents get there in ninety seconds. (That 2.40M is a measured figure from an isolated transcript, not an estimate; it comes with known confounds, noted below.)

Multiply small-per-token by agentic volume and you're back to something real. The International Energy Agency (not an activist blog) projects datacenter electricity roughly doubling to around 945 terawatt-hours by 2030, about the entire electricity demand of Japan, with AI-optimized datacenters more than quadrupling their draw. Each token is a drop. The drops now number in the trillions.

That's the resolution to the doom-vs-dismiss fight. Your one query really is a rounding error. The trillion queries are not. And the part of that trillion you personally control is your automation.

The equation

This is why efficiency and sustainability are the same discipline on this blog, and I mean that literally.

• We cut that /plan run's token cost by 37% with a scoped cache and a throwaway extraction sub-agent; the cache-creation slice alone fell 44%. (Measured before/after on the executor thread, with two honest confounds: some of the 37% is run-to-run variance, and we only measured one thread of a bigger win. The 44% is the piece we can attribute cleanly.)
• Right-sizing the model, per above, is a documented several-fold difference in energy per task.
• Keeping CLAUDE.md a lean index instead of a fat always-loaded doc stops re-paying for context every single turn.
• Killing redundant tool calls and fat payloads stops re-transmitting the same bytes over and over.

Every one of those was pitched here as a way to spend less money. Every one of them is also fewer GPU-seconds, fewer kilowatt-hours, less water through the cooling loop. You don't have to choose between the cheaper path and the greener path. It's one road.

That's the banner line for everything that follows: efficient AI use is an environmental act, not just a budget one.

The honest counterweight

I'd be breaking my own rules if I ended there clean, so here are the edges.

Rebound is real, and I'm not going to pretend I've solved it. If making AI cheaper just makes everyone use more AI, the aggregate footprint can rise even as each token gets more efficient — the Jevons paradox, and a genuine open question at the macro level. What I'll commit to: at the level you control (your architecture, your model choices, your token discipline), efficiency straightforwardly reduces your footprint. The grid-scale rebound is a different lever pulled by different people, and it doesn't let your wasteful pipeline off the hook.

I measured tokens, not liters. Be clear-eyed about what's solid here. Our 37% is a real measured token number. The conversion from tokens to kilowatt-hours to liters of water is external and approximate: it depends on the model, the datacenter's efficiency, and the local grid and water mix, none of which I have Anthropic-specific numbers for. Nobody's published a Google-style metered report for Claude yet. So I'll tell you our tokens dropped 37% and I won't invent a precise liter count to go with it. The direction is certain. The exact magnitude is borrowed from other people's datacenters.

And remember this is all inference, not training. The one-time cost of training a model is a separate, amortized thing. The lever this blog pulls is the recurring inference side: the tokens you emit every day.

The takeaway

• A single query's footprint is tiny; the scary viral energy figures ran about 10× high. Stop feeling guilty about asking a question.
• The footprint that matters is volume × model choice, and both of those live in your automation.
• Right-size the model and cut the wasted tokens. That is the cost optimization and the sustainability optimization: the exact same move, measured two ways.

From here on, every cost post on this blog is also a sustainability post — the cache, the model routing, the lean index, and each gets its own teardown in the weeks ahead. I'm not going to re-argue the footprint math every time. This is the page they'll point back to.

Cheaper is greener. Now go delete some tokens.

FAQ

Does asking an AI model a single question harm the environment?
Barely. Google's metered figures put a median Gemini text prompt at 0.24 watt-hours and 0.26 milliliters of water — about five drops, and roughly one second of a microwave. The older viral figures that claimed each query "drinks a bottle of water" were about 10× too high on energy, per Hannah Ritchie's 2025 analysis. What actually adds up is automation at scale: the pipelines and agents running without you.

What's the single biggest thing I can do to cut AI's footprint?
Right-size the model. A peer-reviewed FAccT study found general-purpose generative models use well over an order of magnitude more energy than task-specific ones doing the equivalent job — 20–30× in their measurements. Picking the smallest model that does the task is the biggest footprint lever most people control, and it's free, because it's the same choice that lowers your bill.

If efficiency just leads to more AI use, does cutting tokens still help?
At the macro level it's genuinely contested — the Jevons paradox says cheaper AI can drive enough extra usage that total footprint climbs anyway. But at the level you control (your architecture, model choices, and token discipline) cutting tokens cuts your footprint directly. Grid-scale rebound is a separate lever pulled by different people, and it doesn't excuse a wasteful pipeline.