Where does AI Rankings get its data?

From open data that its owners publish for reuse: Epoch AI's benchmark data and Epoch Capabilities Index (CC BY 4.0), Arena's human-vote leaderboard dataset (CC BY 4.0), and the open-source LiteLLM price list (MIT licence) for prices and context windows. Every number shows its source and date.

Does AI Rankings run its own tests?

No. We combine and explain other people's open results, and we say clearly which numbers we calculate ourselves, such as the blended price, the best-value models and the consensus score. We do not use Artificial Analysis data and we do not scrape websites.

How often is the data updated?

It is checked daily. Small changes publish automatically. New models and large changes, such as a price change of more than 30 percent, wait for a person to check them. A number older than 45 days is flagged as out of date.

AI Rankings

How AI Rankings Works: Sources, Scores and Prices

AI Rankings compares AI language models on how smart they are, what people prefer, what they cost and how much text they can read. This page explains where every number comes from, what it means and how we work out the few we calculate ourselves.

Data updated 26 Oct 2026

Last reviewed 26 Oct 2026

What this is and who it is for

Choosing an AI model means weighing several things at once: how capable it is, whether people like its answers, what it costs and how much it can read in one go. The numbers are scattered across research sites and price pages, and they are written for specialists.

AI Rankings brings the open ones together, matches the same model across sources, ranks them, and explains each measure in plain words. It is meant for people who use or buy AI at work or for study, and for builders picking a model for a product.

It is a comparison, not an endorsement. A high rank says a model did well on the measure shown. Whether it suits your job is for you to test. Our tool finder and prompt library help with that next step.

Where the data comes from

Three open sources, each used under its own licence. We read each source's published data files or open data feed, never its web pages.

Epoch AI

What we take: The Epoch Capabilities Index (one combined score over many benchmarks), individual benchmark results, release dates and whether a model's weights are open.
Licence: Creative Commons Attribution 4.0 (CC BY 4.0)
Credit: Benchmark data: Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks (CC BY 4.0).
What we change: We match models across sources, rank them and show them in charts. Scores are not changed. Epoch AI does not endorse this site.

Arena (formerly LMArena)

What we take: Human-vote ratings: people compare two anonymous answers and pick the better one. We use the text, coding, web development and vision boards.
Licence: Creative Commons Attribution 4.0 (CC BY 4.0)
Credit: Human-preference ratings: Arena (arena.ai) leaderboard dataset, CC BY 4.0.
What we change: We keep the latest published ratings, match them to models from other sources and rank them. Ratings are not changed. Arena does not endorse this site.

LiteLLM model price list

What we take: List prices per million tokens, context window size and supported features, as recorded by the open-source LiteLLM project from each model maker's own price page.
Licence: MIT licence
Credit: Prices and context windows: LiteLLM model price list (github.com/BerriAI/litellm, MIT licence).
What we change: We use the model maker's own price entry when there is one. Prices change often, so always confirm on the maker's site before you buy.

What we do not use: we do not use data from Artificial Analysis, and we do not scrape any website. Where a site's terms limit copying its data, we leave it out.

What each number means

Epoch Capabilities Index: The Epoch Capabilities Index combines results from many benchmarks into one number. Higher is better. The range shows how sure Epoch is about the score. Because it blends many tests onto one scale, a model does not rise or fall on one benchmark alone.
Arena rating: Arena asks real people to pick the better of two anonymous answers. A model's rating goes up when it wins more of these votes. Higher is better. Arena publishes separate boards. We use four: text (Arena's style-controlled overall board), coding (the coding category of the text board), web development and image understanding (the style-controlled vision board). Style control adjusts for answer length and formatting.
Benchmark scores: A benchmark is a fixed set of questions or tasks. Scores are percentages and higher is better. For each model we show the best result Epoch lists across its effort settings, the same rule for every model. Tests shown: GPQA Diamond, Humanity's Last Exam, SWE-bench Verified, Aider Polyglot, Terminal-Bench, FrontierMath (Tiers 1-3), ARC-AGI-2, SimpleQA Verified, OTIS Mock AIME, SimpleBench, DeepSWE.
Price: Prices are dollars per million tokens. A token is a small piece of a word, so a million tokens is roughly 750,000 words. Output tokens (the answer) usually cost more than input tokens (your question). These are the list prices of the model maker's own API. When a model is only sold by other hosts, as with many open-weights models, we show the median price across hosts and label it. Chat-app subscriptions, free plans, volume discounts and reasoning tokens are not included.
Context window: The context window is how much text a model can read at once, including your question and its answer. Bigger windows suit long documents.
Release date and open weights: Release dates and whether a model's weights can be downloaded come from Epoch AI's model records. A model released only to selected customers is not marked as open.

Numbers we calculate ourselves

These are our own calculations from the open data. They are never presented as a figure published by Epoch AI, Arena or LiteLLM.

Blended price

Blended price assumes three input tokens for every output token: (3 x input price + output price) / 4. It is a typical mix for chat and document work. Your own mix may differ.

Best-value models

Best-value models sit on the cost and intelligence frontier: no other model has a lower or equal blended price together with a higher or equal Epoch Capabilities Index score. Models with no list price, or a price of zero, are left out.

Consensus score

Consensus score: we take every model that has both an Epoch Capabilities Index score and an Arena human-vote text rating, and work out where it sits in each list as a percentile from 0 to 100 (100 = best of the group). The consensus score is the average of those two percentiles. A model needs both scores to get one, and the percentiles are only measured against other models that also have both. It is our own calculation, not a score published by Epoch or Arena. We only show it when at least 8 models have both scores.

Cost of 100 chats

Cost of 100 chats assumes each chat sends 1,000 tokens (about 750 words) and gets 500 tokens (about 375 words) back.

Cost to read a 300-page book

Cost to read a 300-page book assumes about 100,000 input tokens and no reply. A real book varies with page size and language.

How ranks and “too close to call” work

Ranks are worked out among the models we publish that have the score in question, best first. Models with equal scores share a rank. When two models' likely score ranges overlap, or two benchmark results are only a point or two apart, we say “too close to call” instead of naming a winner.

When a page appears in search results

A model page is offered to search engines only when we hold at least 4 kinds of data for it (Epoch score, Arena rating, price, context window, benchmark results) and at least one of them is a quality score. A leaderboard needs at least 8 models, and a single-benchmark leaderboard needs scores for at least 12 models. Tests where several leading models already share the top score are left out because they cannot tell models apart. Thin pages stay out of search results.

How often it updates

The sources are checked daily.
Small changes publish automatically.
New models, and large changes such as a price change of more than 30 percent, a move of more than 5 points on the Epoch index or a changed release date, wait for a person to check them before they go live.
If a source is down, we keep the last good numbers and show their date. Nothing is blanked or guessed.
Any number older than 45 days is flagged as out of date.

Independence

Sponsors never change a score, a rank or an order. Partner Spotlight cards are always labelled Sponsored or Featured and sit below the data.

Rankings come only from the open data and the formulas on this page. Nobody can pay to be added, moved up or removed. If you run an AI product and want to reach people comparing models, see how Partner Spotlight works.

What we do not show, and the limits of the data

No speed or latency numbers. No open source publishes them, and we will not show a number we cannot source.
No reviews or star ratings. We show scores from tests and votes, not opinions.
Tests measure narrow skills. A model can lead on one test and fall behind on your task. Try it on a real piece of your own work.
Tests can leak. If questions end up in training data, a high score may reflect memory. See benchmark contamination. This is why we show several independent sources side by side.
Old tests top out. When most strong models score near the maximum, the test stops separating them.
Setup changes scores. Effort settings, prompts and tools differ between groups, so numbers from different sources are not always comparable.
Human votes reflect the voters. Arena ratings show what the people who vote prefer, which depends on the questions they ask.
Price is not the full cost. We show the standard list price. Some makers charge more for very long inputs, and batch, priority and cached rates differ. Long answers, reasoning and volume discounts also change what you pay. Check the maker's price page before you commit.
Matching names is hard. The same model is spelled differently across sources. We match them carefully and never guess, but a mistake is possible. Tell us if you find one.

Corrections

If a number looks wrong, use Report a mistake and name the page and the number.

We check it against the original source. If our matching or a calculation caused the error, we correct it as soon as we can and the page shows the new value and date. If the source itself looks wrong, we do not edit scores by hand: we can hide the affected model until the source is fixed. Either way we keep the source's credit visible.

Sources and credits

Benchmark data: Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks (CC BY 4.0). Human-preference ratings: Arena (arena.ai) leaderboard dataset, CC BY 4.0. Prices and context windows: LiteLLM model price list (github.com/BerriAI/litellm, MIT licence).

Where this data comes from

Epoch AI: Benchmark data: Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks (CC BY 4.0). We match models across sources, rank them and show them in charts. Scores are not changed. Epoch AI does not endorse this site.
Arena (formerly LMArena): Human-preference ratings: Arena (arena.ai) leaderboard dataset, CC BY 4.0. We keep the latest published ratings, match them to models from other sources and rank them. Ratings are not changed. Arena does not endorse this site.
LiteLLM model price list: Prices and context windows: LiteLLM model price list (github.com/BerriAI/litellm, MIT licence). We use the model maker's own price entry when there is one. Prices change often, so always confirm on the maker's site before you buy.

Newest data on this page: 26 Oct 2026. Prices are list prices and change often, so confirm on the maker's site before you buy. Sponsors never change a score or an order. How we calculate this · Report a mistake