Methodology

How data is collected

• Official pages first. Prices, specs and release dates are read from each provider’s own pricing pages, model documentation and launch posts, and from the public price pages of the channels that sell the model.
• Sources per field. Every value records the page it was read from and the date it was checked, not one source per model. The dot after a number opens that record.
• Checked before publication. The collector publishes changes automatically after the dataset passes validation and the latest replay of stored pages passes. Every change is recorded and can be reverted.
• Unchecked values are marked. A value not yet checked against an official page is shown as unverified, and it never overwrites a verified value.
• No guesses. Missing values are shown as “—”. Open-licence community data may fill a gap as unverified until an official page states the value. We don’t estimate values or compute composite scores.

Verified, partly verified, unverified

Verified

Every price and spec of the model has an official source, and its type’s core specs and at least one priced offering are present. 487 of 1292 models.

Partly verified

Some values have an official source, but other values, core specs or a priced offering are not yet verified. Verified values show a filled dot, the rest a hollow one. 794 models.

Unverified

No value has an official source yet. 11 models.

Cited

A score from a third-party leaderboard, shown with its source and license. We cite it; we didn’t measure it.

Prices

• Prices are stored and shown in the currency the source lists them in, with every tier and condition the source states.
• Headline prices in lists and comparisons come from the provider’s own API at its standard rate: no batch, long-context or off-peak tiers. For image models it is the lowest listed price per output image; for video models the price per second at 1080p where the provider lists it.
• The blended language price is 3 parts input to 1 part output, computed from the listed prices and labeled as such.
• Converted prices, when exchange rates are available, always show the original price and the rate date.

Scores

• Provider-reported scores are benchmark results a model’s own provider published, with the conditions it states (for example effort level or tool use). We list 276 of them across 31 benchmarks. Providers run benchmarks differently, so compare them with care.
• Cited scores come only from leaderboards whose license allows us to show them, and are credited as the license requires. Where a leaderboard ranks several configurations of a model, we show the best one and name it.

Cited sources and attribution

MCP server data

MCP server names, descriptions, versions, endpoints and packages come from the Official MCP Registry, whose data is dedicated to the public domain under CC0 1.0. This data comes from the Official MCP Registry; listing a server here doesn’t imply any endorsement.

Popularity figures are the platforms’ own measurements, refreshed weekly and shown with their date: GitHub stars from the GitHub API, npm downloads from api.npmjs.org, and PyPI downloads from pypistats.org, which serves the PyPI BigQuery dataset (Python Software Foundation, CC BY 4.0; we show the last-month total unchanged). They are facts, not a quality score, and we don’t combine them.

Logos and trademarks

Provider and channel logos are from Lobe Icons by LobeHub, used under the MIT License, shown in one color.

Interface icons are from Lucide, used under the ISC License; some Lucide icons derive from Feather (MIT License, © Cole Bemis).

Logos are trademarks of their respective owners, used only for identification.

Local-run estimates

For open-weight models the site estimates what it takes to run them, from their makers’ own repositories (parameters, config.json, the sizes of the builds they publish) and the hardware makers’ specs. Estimates, never measurements: nobody else’s benchmarks are used.

• Weights: the maker’s own file where it publishes the precision, else parameters × bits per weight (GGUF levels as llama.cpp defines them, K-quant mixes calibrated on makers’ files).
• KV cache: 2 × layers × KV heads × head width × context × bytes per value; latent attention keeps one compressed vector a layer, sliding-window layers keep their window, hybrid models only their full-attention layers. × concurrent requests.
• Overhead: the larger of 0.5 GB and 5% of the weights.
• Usable memory: 90% of a GPU’s; two thirds of Apple unified memory up to 36 GB and three quarters above (macOS’s default); three quarters of other unified memory; system memory less 6 GB.
• Speed: bandwidth ÷ bytes read per token (the active parameters’ weights and the KV cache), at 50–75% efficiency, the middle shown; several GPUs add their bandwidth less 15%.
• With system memory: whole layers on the GPU and the rest in RAM, or a mixture of experts’ experts in RAM and the rest on the GPU; each part read at its own bandwidth. System memory: JEDEC rates (DDR4-3200 dual channel 51.2 GB/s, DDR5-5600 89.6 GB/s, and four or eight channels).
• Grades: runs well at 15 tokens a second or more; runs, slowly below; needs system memory (at 2 tokens a second or more); doesn’t fit.
• Automatic precision: never below 4 bits (or the native precision), leaving 10% of the memory free where it can. Quality first: the highest that runs well in memory, else the highest in memory, else with system memory. Speed first: a 4-bit build.
• First token: 2 × active parameters × prompt length ÷ the device’s dense FP16 compute.

Data freshness

We re-read pricing, lifecycle and changelog pages on every collection run, several times a day, and spec-only pages once a day. The latest full collection run finished on Nov 10, 2026. Each figure links to its source and check date.

• Model prices and specs: 2763 official pages · checked Nov 10, 2026
• Coding plans, clients and MCP proofs: 87 pages · checked Nov 11, 2026
• MCP servers: 142, from the Official MCP Registry · checked Nov 11, 2026
• Agent Skills: 223, from publishers’ own indexes · checked Nov 11, 2026
• News: 54 announcement channels · checked Nov 11, 2026
• Exchange rates: 30 currencies (European Central Bank) · checked Nov 10, 2026

By provider

A provider’s date is when all of its pages were last read and passed our checks. Each figure also shows the date its own source was checked.

Coverage

1292 models from 32 providers, sold through 39 channels.