How Much Water Does ChatGPT Use Per Query in 2026?

By Muninder Adavelli. Scraping, SEO, and proxies for over a decade.

"Every ChatGPT query drinks half a bottle of water" is the sentence I saw quoted everywhere for three years. It is off by roughly two orders of magnitude. The current 2026 answer sits between 0.3 and 6 ml — a thimble, not a bottle.

Where the viral 500 ml ChatGPT water claim came from

The 500 ml figure traces to one peer-reviewed paper: Making AI Less Thirsty by Li, Yang, Islam and Ren at UC Riverside, published in 2023. The team measured GPT-3 running inference in Microsoft's data centers and estimated 10 to 50 ml of water per response of 150–300 output words. The paper's readable summary — "500 ml per 5 to 50 responses" — is the line that went viral, quoted almost everywhere as "500 ml per query." The paper was careful. Its reading was not.

Three things changed since that 2023 water estimate

The 2023 number rested on three inputs. All three moved.

The model. GPT-3 was the paper's subject. GPT-4, GPT-4o, o1, o3 and GPT-5-class models have shipped since. Each release moved per-token compute in a direction the 2023 paper did not — and could not — anticipate. Quoting the old figure for a 2026 chat with GPT-5 is like quoting a 2015 iPhone battery number for the current one.

The efficiency. When Google published Gemini production telemetry in May 2025 — a vendor disclosure with methodology attached, not a marketing figure — it put per-median-prompt energy at 0.24 Wh. Most 2023 academic estimates assumed 2.9 Wh, an order of magnitude higher. Cooling numbers moved the same way. Lawrence Berkeley National Lab, a US Department of Energy research lab, put 2023 US data-center site water use efficiency at 0.36 L per kWh, against industry-average assumptions closer to 1.8 L per kWh in circulation before that. Multiplied through, the water math shrinks by roughly a factor of ten before you touch anything else.

The scope. "500 ml per 5 to 50 responses" is not "500 ml per query." The internet made that translation and never corrected it. That misread alone is a 5 to 50× overstatement, before efficiency or model come into play. And there's a related scope point worth naming: a one-line prompt returning three tokens costs a fraction of what it takes to write a 2,000-token essay. Any single per-query number is averaging over an enormous range of prompt shapes. That is why the honest 2026 answer is a range, not a point.

Stack the three, and a headline that described GPT-3 within a rounding error in 2023 is off by two to three orders of magnitude for GPT-5 in 2026.

What the six best current ChatGPT water studies actually measure

None of what follows are random blog guesses. Two are vendor disclosures (one with methodology, one without), one is a peer-reviewed 2026 study, one is a nonprofit third-party auditor, one is an industry technical writeup with the math shown, and one is a numerate practitioner's independent recalculation.

OpenAI's CEO Sam Altman disclosed 0.32 ml per query in a June 2025 blog post — a vendor claim without published methodology, but from the company running the workload. Google's Gemini production telemetry the same spring showed 0.26 ml per median prompt — a vendor disclosure with methodology attached, which makes it the most auditable number in the set. Sean Goedecke, a senior engineer who writes one of the more numerate technical blogs on modern AI infrastructure, recalculated the per-conversation cost from public inputs and landed at roughly 5 ml. CometAPI, an AI-API tooling company, derived 5.22 ml — from 2.9 Wh × 1.8 L/kWh — with a range of 0.58 to 12.76 depending on cooling assumptions. EcoLogits, a nonprofit auditor of AI environmental impact, measured 6.11 ml for an email-length task on GPT-5.5. And a peer-reviewed September 2026 study, cited in Poynter's fact-check of Altman's numbers, put the range at 0.6 to 17 ml, prompt-length dependent.

Vendor disclosures cluster at the low end. Third-party audits and independent recalculations run higher. That gap is real — vendors have PUE and cooling data outsiders can only estimate — and it's worth respecting rather than papering over.

Why ChatGPT's per-query water use is still a range, not a single number

The number moves with four inputs: data-center location, model, prompt length, and cooling technology. A short prompt in a Nordic data center on closed-loop cooling costs a fraction of a long-context request in an Arizona site with evaporative towers. Shaolei Ren of UC Riverside, one of the authors of the 2023 paper that started this whole conversation, has said this in almost those exact words — resource use varies substantially with location, model, prompt and output length, inference mode, and efficiency. Anyone quoting a single point estimate is averaging over conditions they did not measure. The range is not evasive. It is the honest form of the answer, and it is tightening — Google's May 2025 telemetry disclosure was the first vendor-published methodology worth auditing against, and more of those will follow.

Where AI water use actually adds up — training and image trends, not chat

At a billion queries a day and 5 ml each, ChatGPT's daily inference footprint is roughly 5 million litres — two Olympic pools, a small factory's daily draw, not a city's. Recalculated at the outdated 500 ml assumption, the same billion queries look like 500 million litres a day, and every alarming headline about AI drinking a small country runs from there. The correction changes the political weight of the story more than the raw arithmetic does.

But this is where the per-query framing itself becomes the problem. Training a frontier model runs 7 to 15 million litres per pass, extrapolating from the 700,000 litres the UC Riverside team measured for GPT-3. A single training run consumes more water than a heavy user's decade of prompts. And image generation, when a viral trend hits — Studio Ghibli one week, 80s-retro portraits the next — spikes volume by orders of magnitude while costing more per output than text. The reroll behaviour on image prompts, where one user asks the model to try again twenty times because the first output almost worked, is where the aggregate curve genuinely bends.

I run scrapers hitting LLMs at 2 to 5 million requests a month for an AI visibility tool I'm building. If per-query water were the 500 ml the internet quoted, my monthly draw would exceed a small district's supply. It does not. My daily footprint sits closer to a household's than a neighbourhood's, and the compute bill tracks that. The per-query number has always been a distraction. The real question is how many training passes the grid absorbs, where they land, and how the trend-of-the-week image cycles reshape aggregate demand.

The honest 2026 answer

For a typical ChatGPT text query, roughly 0.3 to 6 ml of water; long context can push 15 ml, image generation more. The number worth arguing about is not per-query. It is per-training-pass, per-trend-cycle, and per-region.