OptionalbaselineCode your app runs instead of a model (a matcher, a rules engine) or alongside it, tested so the board shows whether a model beats it. A baseline is compared like any model but never suggested as live or as the fallback. Give it a label that says what it is.
OptionalcachedUSD per 1M input tokens read from the provider's prompt cache. Unset:
cached input costs inputPer1M, which overstates rather than flatters.
OptionalcacheUSD per 1M input tokens written to the prompt cache (Anthropic bills
these above the input price). Unset: they cost inputPer1M.
OptionaldecisionsAs a decision model: "native" probabilities come from the model itself (a full distribution for choices and scales); "prompted" ones are the model's own estimate written out, with no distribution for choices.
OptionaleffortsReasoning efforts it accepts, lowest first.
OptionalinputUSD per 1M tokens, for estimates and the cost of each answer.
OptionaloutputWho serves it ("anthropic", "openai", …): the fallback must differ.
OptionalretiringWhen the provider retires it, as "YYYY-MM" or "YYYY-MM-DD". A retiring model is never suggested, and a retiring live model gets replaced.
OptionalsourceWhere the price came from, so it can be checked.
What Evalkeeper needs to know about one model: its label, provider, prices, reasoning settings, and retirement date.