Model catalog

Snapshot provenance

This is the catalog every uploaded workspace is priced against. The Thornbury demo is deliberately pinned to a frozen scenario catalog so that catalog updates cannot move its published decision bytes — its figures will not reconcile against this table, and every decision record states the catalog version it used. Each row is a family: a navigation and comparison device that compresses several model versions, so its rate can differ from any single deployment beneath it. The accounting key is a deployment's price schedule — expand a family's endpoints for the per-provider rates a figure is actually computed from. This catalog contains only objectively sourceable facts — published per-token pricing, context windows, rate limits, modalities, deprecation dates. It carries no quality scores of its own, and Bayeto does not rank models it cannot measure. Where an independent benchmark has published a comparison, it is shown as an attributed, directional signal — never a score we assert, and never a number in the arithmetic behind a recommendation. Each family's price is one named deployment's published schedule, and the row states which model version it was captured from. It is not a claim about your traffic: Bayeto groups model versions into families before pricing, so if you run a different version of a family than the one named here, your recommendations are computed at this version's rates. Every cache figure here is one of three things, and they are different answers: a published price; not billed, meaning the provider's own terms establish that no separate charge exists on that leg — a finding Bayeto acts on; or unknown, meaning nobody established it and the engine stands down rather than guessing. Until catalog-2026-07-22-1 the third case was silently written as the input rate, which made an unmeasured cost indistinguishable from a real one and suppressed a rule on 21 of 24 families. Every listing's maker is either identified or declared unstated, and none is inferred — a marketplace's own filing names the creator on one source and the host on another, so both are resolved from curated maps rather than from the model name Nothing is inferred now except where a figure is carried over from a prior snapshot, or a cache leg is derived from a rate the provider publishes as a multiple of its own input price — and either way it is marked estimated. The same rule holds for context windows: an unknown one gates a model out of being a migration target while leaving it fully priceable as a model you already run, and a blank cell cannot carry that meaning. A family with no first-party published price can still be priced at deployment level: the provider range is real even where the family row has no tariff, so per-provider comparison continues where reprice counterfactuals stand down.

Every model, every published price

Source pin: litellm@973329e98641 · Recommendations state the catalog version they were computed against; a new snapshot never rewrites an existing recommendation.