Airside: Model Verification and Crew Invites
Airside now runs a preflight against your endpoint before a model can be filed: every capability you declare is probed live, and the results become the capability badges developers see. Carriers can also pick the upstream API format per model, invite crew by email, file per-model fares, and rename their carrier under review.
A model listing is a set of claims: this endpoint streams, accepts images, calls tools, returns JSON. When one of those claims is wrong, developers find out through a 400 and routing quietly sends their traffic elsewhere. Model verification checks the claims before the listing goes live: Airside runs a preflight against your endpoint and records what actually works.
Preflight Before Filing
When you register a model under Fleet, Airside queues a preflight against your upstream API before the filing can be submitted. You paste a provider API key that is used only by that run and erased when it finishes. Each declared capability gets its own check, so a failure tells you exactly which flag to fix rather than rejecting the whole listing:
| Check | Runs when you declare |
|---|---|
| Basic completion | Always |
| Streaming | Streaming |
| Vision input | Vision |
| Audio input | Audio input |
| Tool calls | Tools |
| JSON output | JSON output |
| Structured JSON | JSON schema output |
| Reasoning | Reasoning |
| Reasoning budget | A reasoning token budget |
| Web search | Web search |
Results update in the dialog as the checks complete. Editing a model after its preflight invalidates the run, so a listing can never ship with checks that were made against different settings, and a running verification cannot be queued twice. Existing listings get a Run verification button, and the verified capabilities appear as badges on the model's provider page on llmgateway.io.
Choose the Upstream API
Not every provider speaks Chat Completions. Each model now declares the upstream API the gateway should use: the carrier default, OpenAI Chat Completions, OpenAI Responses, or Google Vertex generateContent, with the same Vertex OAuth handling the static catalogue uses. The preflight runs through that format, so what passes verification is what serves traffic.
Crew, Fares, and Branding
- Crew invites go out by email to new and existing users. A teammate with an account joins immediately; anyone else joins the first time they sign in with the invited address. Failed sends roll back so owners can retry, and pending invites can be revoked.
- Per-model fares let a single model carry its own traffic discount and landing fee instead of inheriting the carrier's, filed for review like every other fare change.
- Renaming your carrier files a branding change that the LLM Gateway team reviews before it appears on the public providers and models pages.