Every number on the dashboard is calculated the same way, from public data, and checked against laboratory measurement. This page is the working.
The one rule everything else follows. Our figures are accurate to roughly a factor of two to three. So we only ever state a difference when it is at least 5×, comfortably outside that margin. Where two models are closer than that, we say they are too close to call rather than inventing a ranking. This is why the dashboard will not tell you that one model is 20% better than another: we do not know that, and neither does anyone else.
Epoch AI's Capabilities Index, a single score combining 58 separate benchmarks, covering coding, agentic tool use, long-context comprehension, writing and science reasoning. It runs from roughly 55 to 170. We do not compute it ourselves; we take Epoch's published score, and their roster is also what decides which models appear here at all.
Until September 2026 this axis was a single science test, GPQA diamond, scored 0-100. That test is still shown on each model's own page, but it is no longer the capability figure, for two reasons. It gated models off the site entirely, a model Epoch had not run that particular test on simply did not appear, which at one point hid 84 models including several of the newest. And it saturates: at the top of the market almost everything scores in the eighties or nineties, so it stopped telling models apart just where telling them apart matters most.
The change does not reshuffle the rankings. Across the 161 models scored by both, the two measures put models in almost exactly the same order (Spearman rank correlation 0.968, where 1.0 is identical). A small number of older models have no index score at all; we leave their capability blank rather than substituting their score from a different test, because two different scales in one column would put models in the wrong order, which is the one thing a ranking must not do.
Calculated, not metered. No AI company publishes per-answer energy, so there is nothing to look up. We use the EcoLogits model, which works from how much of a model fires per word generated and how many words it writes: a 7,000-word-ish answer for a complex prompt, 300 for everyday chat, on a world-average electricity grid.
What actually drives the number is active parameters × tokens generated. That is worth knowing because it tells you what the figure is sensitive to. Sweeping a model's speed from 2 to 100 tokens per second moves the result by under 5%; parameter count moves it by around 4,000×. Size is the whole story, and speed is a rounding error, which is why the dashboard has no speed axis.
The raw EcoLogits formula assumes a model has its hardware largely to itself, which
overstates real energy use by roughly a factor of two. We correct for that against
46 samples vs ML.ENERGY on B200, batch 32-128, applying a factor of 0.53518.
Laboratory measurement sounds strictly better than a model. It is not, for our purpose, and this is the single most important thing on this page.
The same model, on the same benchmark, measured at different serving batch sizes:
| Requests handled at once | Measured energy per answer |
|---|---|
| 8 | 0.236 Wh |
| 64 | 0.087 Wh |
| 256 | 0.041 Wh |
| 4096 | 0.028 Wh |
An eightfold spread from the serving setup alone, for an identical model doing identical work. And measurement only exists for open-weight models, nobody can measure GPT or Claude or Gemini from outside. So publishing measured figures next to modelled ones would have shown open models as roughly 30× greener than closed ones, purely as an artefact of which ones can be measured. That would have been a ranking of serving configurations wearing the costume of a ranking of models.
So every headline figure here is computed the same way, including for the 12 models we do have measurements for. Measurement is used to calibrate and to check our work, never to replace it.
Cross-checking the calibrated model against measurement on the 12 models present in both sources, at matched answer lengths and matched serving configuration: a bit over half land within 2× of the measured value. The worst overestimate is 3.4×; the worst underestimate is 0.39×.
So: good to roughly a factor of two to three. The error is not random noise, it clusters by model family, which the formula cannot capture.
Permitted. Claims about differences of 5× or more. The headline finding, spreads of 200× and up within a single capability band, sits far outside the error bar.
Forbidden. Ranking adjacent models, or any claim of the form "20% worse than". The dashboard must not imply precision it does not have.
The weakest link is honest to name: the calibration comes from open models between 8B and 405B parameters, and is extrapolated to closed frontier models an order of magnitude larger, whose parameter counts are themselves published guesses spanning 3×. The largest figures on the chart carry the most compounded uncertainty, and are drawn as wide bands rather than points for exactly that reason.
Energy cannot be worked out without knowing how big a model is, and how much of it fires per word. Where a maker has published neither, and nothing in the model's name or Epoch's notes settles it, assuming one architecture over another swings the answer about 30×.
Rather than pick, we carry the full span as an explicit band, flag the model, and bar it from every headline claim. Those models still appear on the chart and in the table, leaving them out would quietly reward the silence. Removing them entirely does not change the headline finding.
The models most people arrive here asking about, the newest flagships from OpenAI, Anthropic, Google and xAI, have no published parameter count from anyone. For a long time they carried no energy figure, which meant this site was silent on exactly the question it exists to answer. 107 models now carry an inferred figure instead.
How it is worked out. Two things predict how big a model is: how capable it is, and when it came out. More capable models are bigger, and for any given level of capability models shrink over time as training improves, the trend in our own data is roughly a four-fold shrink per year, which is large enough that leaving release date out makes the estimate noticeably worse. We fit both effects against the 135 models whose size is actually published, and read the unpublished ones off that fit.
How well it works. We tested it the only honest way: by hiding the size of a model whose size we know, predicting it, and comparing, for every one of those 135 models in turn. The predictions are unbiased overall (the median prediction is 0.99× the true figure, where 1.0 is perfect), and the true value falls inside the band we publish 74% of the time, which is what a band of this width is supposed to do. Fitting on open-weight models and testing on closed ones, and the reverse, showed no meaningful bias either way, which matters, because the makers who publish sizes are not the makers we are predicting for.
What that means for reading them. These are the right order of magnitude, not figures. They are labelled Inferred everywhere they appear, shown as a wide band rather than a number, and, like every other estimate here, barred from headline claims and from any "switch to this instead" recommendation. We would rather publish an honest bracket than leave the biggest models on the market blank, but we are not going to advise anyone on the strength of one.
A remaining 8 of the 285 models here have no size and no capability score, so not even a bracket is possible. That is not a gap in our data so much as a gap in theirs, which is why the dashboard scores each company on how much of its own line-up can be assessed.
Energy is driven by the total text a model generates, including the reasoning it works through and never shows you. 65 of the models here do exactly that. OpenAI, Anthropic and Google each confirm in their own documentation that this text is generated and billed while being returned only as a summary, or not at all.
Our figures count a fixed answer length. So for those 65 models the number shown is an understatement, not an estimate. Each one is marked in the table. Open-weight models are recorded as showing their reasoning because anyone can run them and read it; closed models whose makers have published no position are recorded as undisclosed, never assumed.
| Models with a capability score | 285 |
| With an energy figure | 277 |
| Cross-checked against laboratory measurement | 12 |
| No parameter count published anywhere | 8 |
| Reasoning hidden / shown / undisclosed | 65 / 144 / 59 |
This site tells people a lighter model would do the same job. That is a recommendation, and benchmarks on their own are a thin basis for one: they can be trained against, over-fitted, or simply fail to resemble the work you actually do. So a recommendation has to clear two independent sources before we will make it.
The capability figure itself, a composite over 58 benchmarks, published under a CC BY licence. It folds in SWE-bench Verified, Terminal Bench, Aider Polyglot and DeepSWE for software engineering; OSWorld, The Agent Company and APEX-Agents for tool use; Fiction.LiveBench for long-context comprehension; Lech Mazur Writing for creative writing; and GDPval and METR Time Horizons, which measure performance on real professional work rather than test questions. It covers 93% of the models we have an energy figure for.
Blind head-to-head comparisons voted on by the public, published under CC BY 4.0. This is the only figure on this site that measures whether people actually prefer a model's answers, rather than how it scores on a test. Coverage is thinner, 52%, because it only includes models enough people have voted on.
How they are used together. We name a lighter alternative only when it beats the model on the index, or sits within three points of it when nothing beats it outright. Where both models have been voted on, people preferring the original is a veto. We will not suggest leaving a model with tens of thousands of votes for one that has never been voted on at all. Where the index does not cover a model, we say nothing rather than guess, which is why many pages here name no alternative.
Two limits worth stating. Using the index means inheriting Epoch's judgement about how to weigh 58 benchmarks against each other, which is a real dependency. And even 58 benchmarks plus a public vote do not settle whether a model suits your work. Treat all of this as evidence for a shortlist, not a verdict.
Watt-hours mean little to most people, so each figure also carries two plainer ones: the carbon released generating that answer, and the distance you would have to drive to release the same. The carbon figure leads because it rests on one assumption where the driving distance rests on two, and the driving distance is kept because almost nobody has a feel for what a gram of CO2 is. Every number behind both is stated here rather than buried.
The electricity. 0.45829 kg of CO2-equivalent per kilowatt-hour, the world-average electricity mix published by EcoLogits, which is the same grid its energy model already assumes. This introduces no source we were not already relying on.
The car. 248.5 g of CO2 per kilometre, from the US Environmental Protection Agency, which puts an average passenger vehicle at 400 g per mile. American cars are heavier than the world average, so a smaller car would travel further on the same carbon, every distance shown on this site is therefore shorter than it would otherwise be. Understating is the safe direction. Note that this figure affects only the driving distances, never the carbon figures, which are the electricity alone.
The distances carry two units. Metres and feet below a kilometre, kilometres and miles above it. This is not precision, it is legibility: a reader in the United States does not think in metres and a reader in Britain or Australia does not think in miles, and a comparison nobody can picture is doing no work.
Two honest caveats. Grid carbon varies enormously by country: the same answer generated in Norway and in India differ several-fold, and a world average hides that, exactly as it does in the energy figures themselves. And models whose architecture nobody has published carry neither figure at all, their energy is a wide bracket, and a precise weight of carbon or distance would imply precision we do not have.
All three are public and free to access, and we are grateful for them. Any error in how their data is combined here is ours, not theirs.
Published under CC BY 4.0: use them anywhere, including commercially, as long as you credit The AI Footprint and link back. The underlying dataset is models.json, help yourself, and please cite the three sources above as well as this site.
Visits are counted in aggregate, which pages were opened, from which country, on what kind of device, using Cloudflare Web Analytics. There are no cookies, no accounts, no advertising trackers, and nothing that identifies you or follows you off this site.