Skip to content
Source
Sign inSign up
TanStack/schemas

Vercel reflection/beam: top-level price is now 0 with varies_by_provider, pricing_lost every poll #346

Send to my agent
Closedposthog[bot](GitHub profile, opens in a new tab) opened this issue 1 comment

Observed

Since 2026-10-16 about 18:04 UTC, every Vercel models poll emits pricing_lost (reason: parser_miss) for vercel/reflection/beam. The stored card ($1.30 input / $4.30 output per 1M) is kept, as designed.

Cause (checked against the live source)

  • GET https://ai-gateway.vercel.sh/v1/models now lists reflection/beam with "pricing": {"input": "0", "output": "0", "input_cache_read": "0", "varies_by_provider": true}. The free tag was also added.
  • GET https://ai-gateway.vercel.sh/v1/models/reflection/beam/endpoints shows three routes:
    • baseten: $1.30 / $4.30 per 1M, cache read $0.13. This is the stored card.
    • nebius: $0.25 / $1.00 per 1M.
    • reflection: $0 / $0, tagged free.
  • The top-level price now follows the free route. perToken reads a zero as free, not a billed rate, so tokenCard returns null. The poller keeps the stored card and logs parser_miss.
  • reflection/beam-free is listed separately at 0/0, without varies_by_provider.

Decision needed (not a parser defect)

What should a Vercel row store when varies_by_provider is true and the top-level price is zero, but paid routes also exist?

  1. Treat the stated zero as free and set absent.pricing: 'cleared'. The row's price becomes null, the same as beam-free.
  2. Keep the stored paid card (current behaviour). The alert keeps firing every poll.
  3. Read /endpoints and choose a route price (lowest paid, default, or other). This needs a rule for which route a gateway price means.

The hourly ingest check has not changed code or data. The stored card stays as it is until this is decided.

  1. Update (hourly ingest check, 2026-10-17 ~05:05 UTC)

    • pricing_lost (parser_miss) still fires for vercel/reflection/beam on every poll, most recently at 04:49 UTC. The stored card is kept.
    • /v1/models still lists reflection/beam at input: 0, output: 0 with varies_by_provider: true.
    • New: /v1/models/reflection/beam/endpoints now lists four routes, not three. A togetherai route was added at $0.40 / $1.20 per 1M, cache read $0.04. The other routes are unchanged: baseten $1.30 / $4.30, nebius $0.25 / $1.00, reflection $0 / $0.
    • This is still blocked on the decision above. No PR was opened.
  2. tombeckenham closed this

Sign in to join this conversation.