Third-party scores, model by model.
The published boards Vitna carries, each one its own scale with its own source, and none of them added together. A row is a door to the board's own page.
37 boards · 192 models scored · 8 sources·retrieved sha256 · 0bb53a77713f
Five more publishers are registered and refused: two publish scores nobody licensed, two license them for non-commercial use only, and one publishes none a reader can enumerate. Read , each board on its own publisher's clock.
These six are never added together, because 54 columns measure 54 different things on scales that do not convert. The other way to put them on one axis is to fit one, which is a capability per run and a difficulty per benchmark rather than an average: 319 ranked runs over 46 columns, with the evidence behind each one printed beside it. That scale is Vitna's arithmetic over these publishers' figures, and it is the only number on this page nobody else produced. Its ordering is checked against the published index above, which is carried whole so the check can be read rather than taken on trust; the numbers on the two scales do not compare.
The notice each licence asks for.
Epoch AI Benchmarking Hub data, by Epoch AI, under CC BY 4.0.
MTEB results, dedicated to the public domain under CC0 1.0. Attribution is not required and is given anyway.
LiveBench, Apache-2.0. Results produced by the LiveBench maintainers on questions refreshed monthly.
HELM, Apache-2.0, by the Stanford Center for Research on Foundation Models.
Vectara hallucination leaderboard, Apache-2.0, by Vectara. Scores produced by Vectara's own evaluation of model outputs.
Aider polyglot benchmark, Apache-2.0, by the Aider project. Results are submitted by contributors rather than produced by the publisher.
Vitna ran none of these evaluations and adjusted none of these figures.
Vitna has run one benchmark itself, of its own router against Jev Router on MT-Bench. It is not a board here, because a measurement a party makes of its own product is a different kind of figure from these; it is with the router, every judge's column printed: the Switchyard benchmark.
Best value is the board's own score divided by the blended OpenRouter list price, three input tokens to one output. A model's listed variants share its scores and are not ranked twice. For the same models priced, see the catalog and the rankings.
OpenRouter's own task benchmarks.
OpenRouter runs task benchmarks of its own, among them GPQA Diamond, tau-bench and a search suite, and publishes the results through its Data API, which needs a key Vitna does not hold. They join this page when they can be read and dated the way these two sources are. The boards here are the ones the public model list carries, nothing more.
Source: OpenRouter (openrouter.ai/models), as of 17 Nov 16:29 UTC. sha256 0bb53a77713f over 853,270 bytes.
Catalog: OpenRouter's public model list.
Intelligence, coding and agentic indexes: Artificial Analysis, as published by OpenRouter.
Elo, win rate and rank: Design Arena, as published by OpenRouter.
Vitna ran none of these evaluations and adjusted none of these figures.
A board appears when at least five ranked models stand on it. 186 listings carry a Design Arena row; the boards under the floor are listed on each model's own page.
