Every id, on a card.
The models this gateway will actually serve, each on its own plate: lane, context window, the price per million tokens you are billed, and the privacy tier it genuinely runs under. Read live from the running gateway — not a brochure.
Everything below is what GET /v1/models returned just now. A model pinned to an upstream this deployment has not configured is omitted by the gateway, and this page shows the omission rather than backfilling it from a hardcoded list. Prices already include Sable’s margin, so the figure on the card is the figure on the receipt.
Lanes are filters, not buckets — a confidential model is also open-weight, so the counts deliberately do not sum to the total. Price meters are log-scaled against the dearest model in the catalog (it spans $0.012 to $60, so a linear bar would hide most of it); context meters are plain proportional. Length is a real number either way — it is never a rating.
Read the tier, not the vibe.
The Private lane routes caller → Sable → Venice → vendor, so the vendor never learns who you are. It hides who, not what: the prompt text is still readable at the far end, and these models are standard tier for exactly that reason.
Confidential execution means running inside an enclave Sable attests, which needs downloadable weights. Every confidential id here is open-weight, and no closed-weight model is shown under that tier — the gateway enforces it, this page only prints it.
sable, sable-fast and sable-max are stable names pointed at strong engines. That is the product; the engine is disclosed on every receipt, and repointing it is a catalog edit rather than a breaking change.
Nothing on a card is awarded. Every tag is a field the gateway returned; every meter is a real price or a real context window on a stated log scale. There is no quality ranking here because Sable cannot substantiate one.