LLM Gateway #402
Send to my agentAdd LLM Gateway (llmgateway) as a monitored provider.
Sources (both public, no key)#
- Schemas:
https://api.llmgateway.io/openapi.json(OpenAPI 3.0: chat, embeddings, images, audio, video, moderation, rerank, …). - Catalog:
https://api.llmgateway.io/v1/models— 298 models, each withproviders[](627 routes), each route with its ownpricing,max_output,reasoning_efforts.
Pricing#
A bare id (gpt-oss-120b) is routed by a live weighted score (price 0.6, uptime 0.5, throughput 0.3, latency 0.075 — docs.llmgateway.io/features/routing), so its billed price is not fixed. The listing's top-level pricing is undocumented (usually the cheapest route) and is not used.
- Bare rows (
gpt-oss-120b): a rate card only when the model has one route, or every route quotes the same price; otherwise null. - Pinned rows (
groq/gpt-oss-120b): one row per route.provider/modelis LLM Gateway's documented syntax to force one provider; the row carries that route's exact price. - Billed at the provider's own rate; LLM Gateway's 5% fee is on credit purchases, not per request, so it is not in the cards.
- No API change: both are ordinary model rows. Multi-price-per-row is out of scope (follow-up if wanted).
sameAs#
Links come from the route LLM Gateway states (providers[].providerId + externalId), never from family (loose: e.g. "Qwen3.8 27B"). Pinned rows link to their route; bare rows link only when they have exactly one route; multi-route bare rows stay unlinked. Route provider names map to ours via provider_model_namespaces (e.g. aws-bedrock → amazon-bedrock); ambiguous names (bytedance, alibaba) are left unmapped.
Done when#
- Adapter + tests, provider seeded, schemas sync from the OpenAPI doc, models poll with the rules above.
docs/source-silent/llmgateway.mdlists facts the listing does not publish.
tombeckenham closed this
Sign in to join this conversation.