AI Water Cost Calculator: Step-by-Step Water and Cost Estimates
Reviewed by the OnlineFree.app team · Updated
Key points
- The AI Water Cost Calculator turns model, total tokens and request count into cost, litres, kWh and grams of CO₂e.
- All computation runs in the browser from a static coefficient table, with no signup and no uploaded data.
- Water is derived as energy in kWh multiplied by a water usage effectiveness value.
- Results scale roughly linearly with tokens, so 100M tokens costs about 100× a 1M-token run.
- Request count only adds fixed per-call overhead, which matters most for many short calls.
What the AI Water Cost Calculator answers
The AI Water Cost Calculator estimates the money, water, energy and carbon behind an LLM workload from three inputs: the model, the total token count and the number of requests. Open AI Water Cost Calculator, choose a model, type your workload size, and the result card fills in immediately.
Everything runs in your browser against a static per-model coefficient table. There is no account, no upload and no server-side processing of your figures, which is why it stays fast even on a phone. The output is deliberately an order-of-magnitude estimate — useful for planning, sprint reviews and journalism, not a replacement for a provider invoice.
It exists to answer a question a billing dashboard never shows you: what did that prompt actually cost, in litres and kilowatt-hours rather than just dollars.
How do you run an estimate step by step?
Start with the Model dropdown. The options are GPT-4o, GPT-4o mini, GPT-3.5 Turbo, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3 70B (self-hosted) and Generic average LLM, with GPT-4o as the default. That single choice selects the coefficient row used for price, energy per token, water per kWh and carbon intensity.
Next, enter Total tokens — prompt plus completion, summed across the whole workload, not per request. Quick-set chips jump you to 1K, 1M or 100M. Then enter the number of requests; this only adds fixed per-call overhead such as network handshake and cold start on top of the token cost. The 'Use typical: 1M tokens / 10k requests' preset link fills both fields at once if you just want a feel for the scale.
Read the result card top-down: two hero tiles for Cost in USD and Water in litres, then two smaller readouts for energy in kWh and carbon in grams of CO₂e. Nothing below the tiles is a new input — all four values come from the same three fields. The water tile pairs its litres with a human comparison line, such as a multiple of 500 ml bottles, so you can sanity-check the magnitude at a glance.
What each of the four numbers means
Cost is a blended token price multiplied by token count, plus a small per-request handling term. It approximates what a hosted API would charge at list price, so caching, batch discounts or committed-use rates can pull a real invoice lower.
Energy comes from per-token energy coefficients. Water is that kWh figure multiplied by a water usage effectiveness (WUE) value, split into on-site cooling and off-site power-plant water, and carbon is kWh times an assumed grid carbon intensity. The EPA WaterSense programme applies the same 'use per unit of function' logic when it defines water efficiency, and the IEA's data centre analysis is a solid background read on how energy and water demand at data centres have moved.
Model choice matters more for the cost tile than for the water tile. The Llama 3 70B (self-hosted) row is the one to use when you run your own GPUs: it assumes your datacentre's WUE and grid mix, which can swing litres and grams far more than the difference between two hosted models.
Worked example: 1M versus 100M tokens
The defaults build good intuition. At 1M tokens across 1,000 requests you are at 1,000 tokens per request — a short chat exchange — so per-call overhead barely registers and the token term dominates all four outputs.
Push to 100M tokens with the same 1,000 requests and every figure scales roughly 100×, because the computation is linear in tokens. That is why a 100M-token experiment is a budget conversation and a 1M-token one usually is not. The Cost tile adapts its formatting to match, showing sub-cent values like $0.0042 for tiny runs and four-figure totals for large ones.
Now hold tokens at 100M but raise requests to 1,000,000. You are down to 100 tokens per request, and fixed per-call overhead stops being noise — the request field is what tells the calculator about that. If your product makes many tiny calls rather than a few long ones, this is the configuration worth testing before you ship.
Limits, and how to report the result
The coefficient table is static and deliberately coarse. It ignores model training and fine-tuning, embodied hardware and water, network transit, retries and failed calls, and any savings from caching or batching. Real WUE and grid carbon intensity vary by region, season and hour, so one coefficient per model cannot be right everywhere.
If the number is going into a public post, round it, state the model, token count and request count you entered, and say plainly that it is an order-of-magnitude estimate from a static table. Where a provider publishes its own PUE, WUE or emissions figures, prefer those and use the estimator for the workloads nobody has measured.
For internal sanity checks it is more than good enough, and it is free — the whole thing runs on OnlineFree.app next to other browser-side calculators, like the Timezone Overlap Finder for distributed teams. If a figure is going to change a contract, a disclosure or a headline, verify it against provider data first.
Frequently asked questions
Is the AI Water Cost Calculator free, and does it store my inputs?
Yes, it is free and nothing is uploaded. The AI Water Cost Calculator runs entirely in your browser from a static per-model coefficient table, so there is no account, no signup and no server-side record of the model, token count or request count you entered. Reloading the page simply resets the fields to the defaults of GPT-4o, 1,000,000 tokens and 1,000 requests.
How accurate is the AI Water Cost Calculator's water estimate?
It is an order-of-magnitude estimate, not a measurement. The AI Water Cost Calculator multiplies energy per token by a fixed water usage effectiveness value, which cannot capture regional grid mix, seasonal cooling or a specific datacentre's design. Treat small differences between two runs as noise and expect the litres figure to sit within a broad range rather than being exact.
Does the number of requests change the AI Water Cost Calculator's result much?
Only when individual calls are small. The AI Water Cost Calculator derives almost everything from total tokens and adds the request count purely as fixed per-call overhead for network and cold start. At 1,000 tokens per request that overhead is minor; at 100 tokens per request it becomes a meaningful share of the total.
Which model should I pick if I do not know what serves my traffic?
Choose Generic average LLM for a rough mid-range figure. Pick GPT-4o if you use a hosted frontier model and want the default row, and select Llama 3 70B (self-hosted) only when you run the GPUs yourself, because that row assumes your own datacentre's water and grid intensity rather than a provider's.
Can I cite the AI Water Cost Calculator in a sustainability report?
It is not suitable as a primary disclosure figure. The AI Water Cost Calculator is a static-coefficient estimator that excludes training, embodied hardware and network costs, and it cannot reflect a specific facility's measured performance. Use it for internal sizing and communication, and cite provider or regulator-published data for formal reporting.