Skip to main content
Models & Routing

Supported Models & the Relay Catalog

How the curated model catalog is built, kept current, and priced — and where Workers AI serves a model directly.

The catalog is curated, not the full upstream list

Relay does not expose every model a reference pricing feed knows about. A model is servable only when it has both:

  1. An enabled entry in the curated Relay model catalog (admin-managed), and
  2. An active rate card (a current, priced entry) for it.

A catalog entry with no active pricing is dropped rather than ever being quoted at $0.

GET /relay/v1/models

returns the current curated, priced catalog, with exactly one model flagged as the default — the one VG Code uses when you don't specify --model.

Where completions actually run

Relay serves a request from one of two places, transparently to you — the price you're quoted is the same either way:

  • Cloudflare Workers AI, when the model has a mapped Workers AI equivalent. Served directly off Vibgrate's own Cloudflare account, never leaving Cloudflare's network.
  • The upstream provider, proxied through Relay, for models with no Workers AI mapping — most proprietary frontier models (Claude, GPT, Gemini, etc.).

Which backend served a given request has no effect on billing: both settle against the same published rate card for that model.

Keeping prices current

Relay's rate cards are kept current by a daily sync job that compares against a reference pricing feed of provider list prices (not a resale relationship — Relay still calls providers directly or via its own routing; the feed only tells the sync job what current list price is).

  • Price decreases and brand-new models auto-apply — a customer is never disadvantaged by a sync landing unreviewed.
  • Price increases are held for admin review and never go live silently. Every approved increase notifies every workspace that used that model in the trailing 30 days, both in-app and by email.
  • Every rate is versioned, never mutated in place. A rate change closes the old entry and opens a new one, so a historical request can always be re-priced against the exact rate that was active when it ran — nothing is ever silently re-priced after the fact.

Bring your own model access

If a model you need isn't in the curated catalog, or you'd rather not spend Relay credit at all, BYOK passthrough lets you supply your own provider key and route through Relay for governance and audit only, at no Relay token cost.

Related Documentation