Why does a reasoning model cost so much more?

Energy scales with the tokens generated, one at a time. “Thinking” models emit far more of them, some use tens of times more tokens to answer the same question, so they can cost 10-50× more per task even on the same hardware.

Why can the same model vary so much per query?

Input length, output length, batch size, GPU type, utilisation, the serving engine, and the data-center overhead all swing the number. Measured energy for one model class spans up to ~25× across tasks, which is exactly why we lead with a range, not a single figure.

What’s the difference between location-based and market-based carbon?

Location-based uses the physical grid’s average intensity where the compute runs. Market-based credits the clean energy the operator contractually buys. We headline location-based as the honest physical footprint and show market-based separately, since it often reads near-zero for hyperscalers because of annual renewable matching, which isn’t the same as hour-by-hour clean power.

Why don’t you use the “bottle of water per query” figure?

That number came from a 2023 GPT-3-era estimate that was per a whole conversation in one specific data center. Press coverage compressed it into “one query.” Modern measured per-query water is roughly a thousand times smaller, so the bottle figure is superseded and misleading.

On-site vs grid water: which do you count?

Both, as separate terms. On-site water is what evaporates in data-center cooling; grid water is embedded in generating the electricity, and it dominates. We report consumption (water not returned), and weight it by the region’s baseline water stress.

How accurate are the closed models like GPT and Claude?

Less accurate, and we say so. OpenAI, Anthropic and Google don’t publish the parameter counts or throughput the model needs, so those figures are inferred from comparable open-weight models and flagged “estimated” with a wider uncertainty band. Google’s single disclosed per-prompt figure is the one closed-model anchor we can calibrate against.

Does this include training the model?

No, this is inference only (the cost of running the model), which is what scales with your usage. Training is a large one-time cost amortised across billions of queries; for high-volume inference it adds a small per-query amount we may fold in later.

What about images and video?

They’re much heavier and we keep them out of this first version. Generating an image can cost tens of times a text reply, and a video clip orders of magnitude more, so we don’t extrapolate text coefficients to them.

How often are the coefficients updated?

Whenever primary sources publish new measurements: leaderboard revisions and provider disclosures. Efficiency moves fast (one provider reported a 33× per-prompt energy drop in a single year), so every figure carries a source date and superseded numbers are flagged.

How should I offset this footprint?

Reduce first, then neutralize the rest with high-integrity credits, ideally durable carbon removal. That’s exactly what the 1clickimpact API automates: one call captures the carbon (or plants the trees) for a given amount, so you can wire offsetting straight into your product per transaction.

Back to the calculator

FAQ

Questions, answered honestly

The footprint of AI is genuinely uncertain, and a lot of the numbers online are wrong. Here’s the straight version.

Providers rarely publish per-query energy, and the true number depends on your exact prompt, the data center, the batch size, and the grid at that moment. We use the best measured public data and show a range with a confidence flag, treat outputs as informed order-of-magnitude estimates, not meter readings.

Neutralize your AI’s footprint

Capture the carbon in one API call, per transaction, automatically.

Start offsetting →