About and sources

Who maintains it, and corrections

ModelAtlas is maintained through its GitHub repository, FilledVaccum/Model-Sub. The code, the curated data and every change to them are in its history.

Corrections. Every page has a “Report incorrect data” link at the bottom. It opens a GitHub issue that already names the page; add the correct value and a link to where you checked it. A correction goes into the curated data with that source link and the date it was checked. Raw source data is never edited by hand. The fix appears after the next rebuild. Without a source, a report is noted but the value is not changed.

Methodology

• Sources. Catalogue and prices come from public datasets (models.dev, LiteLLM, genai-prices) and from provider and gateway listings. Hugging Face and Epoch AI add licences, parameters and release facts. The table below lists every source and when it was last read. Hand-checked values carry a source URL and the date they were checked. Each price is taken whole from one source, and the endpoints table names that source and the date it was read.
• Which prices set a model's headline price. The price range, the median and the “cheapest” label use only endpoints that are a direct provider (not a gateway), billed per token (not a subscription), a paid price (not a free tier, batch, plan-included, local or unconfirmed $0 price), the standard service tier, not deprecated, the default region (not a regional or data-zone variant), and not flagged as a price outlier or by a sanity check. The provider must also be vetted: the model creator's own API, a major cloud, or a listed established inference provider. Any other reseller counts only inside a sanity band: every priced side at least 0.6× the creator's own price, or, when the creator has no price, between 0.2× and 5× the median of the comparable endpoints. Outside the band it stays listed with an “unverified / unusually cheap” tag, but it never sets the headline. “Cheapest” is the lowest blended cost at 3 input tokens to 1 output token.
• How plans are checked. Each plan is copied by hand from the provider's own pricing page with the page URL, a quote and the date it was checked. A plan last checked more than 60 days before the build is marked “may be out of date”. A quote taken from a summary of the page, not copied word for word, is tagged “paraphrased from source”.
• Update schedule. The pipeline is scheduled to read every source again and rebuild the site every 6 hours (01:23, 07:23, 13:23 and 19:23 UTC). The data on this build was generated 2026-10-20 03:37 UTC. Plans and other hand-checked values change only when someone checks them again.

Attribution and licences

models.dev, LiteLLM and genai-prices are MIT licensed. Their copyright notices apply to the data derived from them.

Epoch AI, 'Data on AI models'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/data/ai-models'. CC BY 4.0

Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks/use-this-data' [online resource]. Accessed 19 Oct 2026.

Provider logos from models.dev (MIT) or supplied by us; logos are trademarks of their owners and are used only to identify them.

How to read the data

• Unknown is not no. A dotted dash means no source reports the value. A cross means a source says it is not supported. A plain dash means nothing is listed. Zero-data-retention is only shown as supported when a source confirms it.
• One source per price. A price is taken whole from one source and never mixed across sources.
• Curated facts are the only hand-entered values. Each one carries a source URL and the date it was checked; where a curated value overrides a feed, the feed's value is kept as the losing value.
• Price ranges, not single prices. A model's price is shown as the min–max across comparable direct providers, with the median available. Comparable means headline-eligible, as described under Methodology. Endpoints flagged as price outliers stay in the table with a plain note such as “6.2× median”, but are kept out of the range. Quantization classes can be compared separately when at least four endpoints share one.
• Subscription plans (20 providers, such as coding plans) list $0 or API-equivalent rates for a flat-fee quota. They are labelled “subscription plan: not per-token”, left out of ranges and never called cheapest.
• Providers vs gateways. Gateways resell other providers' capacity. They never count towards “providers”, but their list prices are visible: a model with no direct provider shows “via <gateway> $x”, and the model list has an “Include gateways” switch. Gateway rows show the route type, who they route to and the difference from direct prices where known. An effective price (with fees) appears only when the gateway's fees were verified; otherwise the row says “list price; fees not verified”.
• Prices are USD per 1M tokens, or per image, minute, page, query or 1M characters for models that are not priced in tokens; units are never mixed. When a prompt passes a context-tier threshold, that tier's prices replace the standard ones for the whole request. Batch, flex and priority service tiers overlay the standard price. Time-of-day prices (for example off-peak windows) are shown in UTC.
• Model identity. Provider-specific IDs are matched to one canonical model by rules, not fuzzy matching. Each match carries a confidence; below 0.8 the model is hidden from the default list and tagged unverified, meaning the identity was inferred, not stated by a source. Dated snapshots are folded into the model's endpoint instead of being separate models.

Plans, pay-as-you-go and price kinds

• Pay-as-you-go vs plans. An API endpoint charges per token (or per image, minute and so on); you pay for what you use. A subscription plan charges a flat fee, usually monthly, for an allowance of usage inside the provider's own apps or tools, and some include API usage. The two are different products. We never turn a plan into a price per token, never rank a plan against an endpoint, and never convert between currencies without a stated rate.
• Plans are curated by hand. Each plan comes from the provider's own pricing page and carries a source link, an exact quote and the date it was checked. Limits are shown as the provider states them. If a figure is not published it is left blank, not estimated. Quotes taken from a summary of the page are tagged “paraphrased from source”. Coverage is partial: a provider with no plans listed has not been researched yet, which is different from having none.
• Staleness rule. A plan last checked more than 60 days before the site was built is marked “may be out of date”. The marker never hides the data; it tells you to check the source link first.
• Price kinds. A $0 in a source can mean several things, so an endpoint is never shown at a bare $0. Paid shows the price. Free tier means the provider offers free use within limits. Included in plan means it is billed through the provider's subscription plans. Runs locally means it executes on your own machine. Listed at $0 (unverified) means the source says $0 but nothing confirms it is actually free. An endpoint with no pricing keeps the unknown mark. Only paid endpoints enter price ranges, “vs cheapest” and “cheapest” labels. Evidence for the non-obvious kinds is curated with a source URL.