Skip to content

About

About the data

We collect model prices from provider listings and benchmark scores from published results. This page explains which price we show when several providers sell the same model and how we compare scores.

Each score has its own explainer: llminfo IQ for the capability score, Coding IQ for coding capability on its own scale, Value for money for what a model costs per task against its results, and the Yap score for how many tokens a model uses per task against models with the same results.

Where the data comes from, and how fresh it is

Everything on this site is collected automatically from provider APIs and pricing pages, model marketplaces and public benchmark leaderboards. That's a few hundred sources. We run a data refresh every six hours. “Updated” shows when the last refresh finished, and “IQ calculated” shows when we last calculated llminfo IQ.

Collection is automated, and sources disagree, lag and change format. Treat every number here as a well-sourced starting point, not a quote. Always check with the provider before you commit to a price.

Release dates mean one thing across the site: the earliest date the model became publicly available, by API or as open weights. Most come from the maker's own listing. Where the maker gives none, we research some dates one by one against public announcements and reporting. For open-weights models, the date can come from the model's Hugging Face repository. A missing date means we haven't verified one, not that the model is unreleased.

How prices are compared

Each table compares prices in the unit its section states: $ per 1M tokens for text, per image, per second of video, per minute of transcribed audio, and per 1M characters of generated speech. Document reading (OCR) shows two units side by side, per page and per 1M tokens, because a page costs a different number of tokens on each model. Providers quote prices in their own currencies. About five hundred live listings are priced in CNY, EUR, IDR or INR. We convert these to USD when the page is shown, for sorting and comparison. Converted figures are marked with ≈, and the provider's own price is kept. A price stated only in a provider's own credits isn't a currency. It can't be compared in USD, so we treat those listings as unpriced rather than convert them wrongly.

“Cheapest” figures are cautious on purpose. They compare prices in USD, and a figure converted from another currency is marked ≈. They leave out listings flagged as suspicious undercuts. A reseller pricing a closed model far below the maker's own price can't plausibly be real. A flagged price still shows on the model page, with an “undercut?” marker. It just never sets a cheapest figure.

Listings with no stated price are hidden from the price tables rather than shown as blanks. Model pages count them in a note instead. A listing that's included in a subscription and also sold pay-as-you-go is shown at its pay-as-you-go price, with a “sub” marker.

Who sells each listing, and who runs it

Each listing names the billing seller. When a different, known party runs the model, that operator appears on a second line. If the operator is not stated, a marketplace may still name a different selected provider. A missing operator or selected provider means it is not stated, not that none exists.

Listings sort and filter under the seller. Listings sold and run by the model's maker, and documented listings from the maker's official partner, are the first-party ways to buy a model. Other hosts are never presented as the maker.

Comparisons use the maker's own price where the maker sells the model and runs it. Otherwise they use a documented listing from the maker's official partner. If neither exists, they use the market median: the middle of the model's comparable prices, marketplace listings included. A marketplace quote never counts as the maker's or the partner's price.

Provider pages distinguish services, not just companies. Z.ai and BigModel have separate pages even though both belong to Zhipu, because they're different storefronts with different currencies, terms and catalogs. Each page names the parent company and links its sibling services.

Absence means “not stated”, never “no”

This is the most important rule for reading the site. A missing benchmark score, price, knowledge cutoff, open-weights badge or data-policy statement means no source stated one. It doesn't mean “no” or zero. Where a negative is shown (“Says it keeps data”, “Proprietary”), a source stated it explicitly.

Data-policy chips and notes quote the provider's own published policies, tagged with the source they came from. We re-read the live pages at every refresh, so a statement can disappear just because a provider reworded a page. Tier differences are real, and we keep them per plan. A provider's free tier may state that it trains on data while its enterprise tier states nothing about it. So plan pages are the authoritative view for subscriptions. None of this is legal advice. Check with the provider before you rely on it.

One model, many spellings

Providers spell the same model many ways: vendor prefixes, dated snapshots, :thinking variants, TEE/ and quantization tags. We fold these into one model, so one model page gathers every provider's listings. The line under a row shows the provider's exact model id plus serving details (variant · quantization · regions).

“Open weights” means the model's weights are publicly downloadable. Either a source says so, or a public Hugging Face repository shows it. It says nothing about the licence: open weights is not the same claim as open source. “TEE” marks listings served from an attested trusted-execution enclave that the provider can't read into, according to the provider's listing.

Flags are stated facts, not judgments. A flag next to a model's name is the maker's stated headquarters country, with the citation on the model page. Flags on a listing row are the countries the provider's own stated serving regions are in. We map cloud regions to countries using the cloud vendors' published region documentation. No flag means no source stated it. We never guess. A flag says nothing about jurisdiction, applicable law or data handling beyond the cited statement itself.

Benchmarks and llminfo IQ

Benchmark scores come from public leaderboards and provider claims. Matching scores to models is fuzzy at the margins, because versions, dated snapshots and test setups differ between boards. Each score links its source where we know it. Scores from different benchmarks aren't directly comparable. Each table column is its own scale.

Scores are shown in one common form, which can differ from how a source's own table prints them. Fractional results (0.873) appear as percentages on a 0–100 scale. Each benchmark is listed under one canonical name, even where sources spell it differently. The figure is the source's; only its presentation changes. A score marked self-reported is the model developer's own claim, relayed by a source that didn't verify it.

Boards named “llminfo Elo” aren't any source's own leaderboard. We compute those ratings ourselves (Bradley-Terry, on the usual Elo scale) from human pairwise-vote data that a third party publishes under an open licence. Wherever the ratings appear, we name who published the vote data and its licence. Each board has its own scale, set so that its rated models average 1000. A rating compares only with others on the same board, never across boards or with another site's Elo. The ± figure is a 95% interval from the fit, and it understates the true uncertainty.

llminfo IQ condenses those results into one comparable capability score per model type. 100 is average, and each 15 points is one standard deviation. Each score shows how much data sits behind it. It has its own explainer, covering the method, the live weight table and what the score is not.

Model families

Each family groups the generations and sizes in one model line, such as Claude, Qwen or Llama, into a single record you can browse. We build families automatically from model names and providers' model lists, and they change with the data. Groupings can split, merge or be renamed as those lists change. Old family links redirect where the successor is clear. A model with no recognisable line belongs to no family.

Model tables mark a model “In the news” when a recent headline actually names it. Model pages show those headlines, and a few notable links, next to the overview. Provider pages show them next to the listings. They appear only when there are any. Absence means none are listed, never that nobody wrote about it.

Something looks wrong?

A missing provider usually means we don't collect it yet. A price that doesn't match your invoice usually means negotiated or regional pricing, taxes, or a source that lags the provider's page. You can report a correction for review with the link below.