Admin Gateway model picker offers IDs Vercel AI Gateway cannot serve (alibaba/qwen3.8-flash-next)

Describe the bug

The admin "Add Gateway model" picker can offer alibaba/qwen3.8-flash-next (label Qwen 3.8 Flash Next, with Vercel-style USD pricing filled in). Saving or using that model against Vercel AI Gateway then fails with model_not_found / "The model does not exist".

The picker is treated as a callable catalog. In this case it is not.

To Reproduce

1. Open Admin -> AI setting -> Add Gateway model.
2. Search for qwen3.8-flash-next / Qwen 3.8 Flash Next.
3. Select alibaba/qwen3.8-flash-next. The form fills:

Model ID: alibaba/qwen3.8-flash-next
Display name: Qwen 3.8 Flash Next
Input: 0.00000012
Output: 0.0000004
Cache Read: 0.00000001
Cache Write: 0.00000375

1. Add the model and use it (or test it) through the configured AI Gateway.

Expected behavior

The picker should only offer models the configured Gateway can actually serve. A model that returns model_not_found should not appear as a selectable catalog entry, or Add should be blocked until a live test succeeds.

Actual behavior

The ID is selectable in the admin UI. Calling the Gateway returns:

{
  "error": {
    "message": "The model 'alibaba/qwen3.8-flash-next' does not exist",
    "type": "model_not_found",
    "code": "model_not_found"
  }
}

Observed against the public Gateway catalog on 2026-09-11:

• GET https://ai-gateway.vercel.sh/v1/models (371 models): includes alibaba/qwen3.8-flash, does not include alibaba/qwen3.8-flash-next.
• GET https://ai-gateway.vercel.sh/v1/models/alibaba/qwen3.8-flash -> 200
• GET https://ai-gateway.vercel.sh/v1/models/alibaba/qwen3.8-flash-next -> 404 model_not_found
• Vercel still has a marketing page for the ID: https://vercel.com/ai-gateway/models/qwen3.8-flash-next (provider listed as runinfra, not Alibaba Cloud)

Client

• OS: macOS
• Browser: Chrome
• Version: 153

Platform

teable-deployment (Helm), app release.2026-09-11T07-36-49Z.3059

Additional context

Teable fetches the live list from https://ai-gateway.vercel.sh/v1/models in apps/nestjs-backend/src/features/ai/ai-gateway-models.service.ts and caches it in Redis for 1 hour. The admin Add dialog copies that list into the search box (GatewayModelsStep / AddModelDialog). Add only requires id and label; a successful model test is not required.

So a stale cache, a catalog entry that is listed but not routed, or a marketing-only ID can all become a selectable "supported" model. Qwen3.8-Flash-Next is an open-weight / RunInfra-hosted preview; the production Gateway ID that does exist is alibaba/qwen3.8-flash.

Suggested fix:

1. Treat GET /v1/models (or a per-id lookup) as the source of truth for selectable IDs, not the marketing page.
2. Drop models that 404 on per-id lookup before showing them in the picker.
3. Require a successful Gateway test before Add, or at least block Add when the live catalog does not contain the ID.
4. Revalidate the cached catalog when the Add dialog opens, instead of serving a 1-hour snapshot as if it were still callable.