What happens#
The OpenAI models poll has emitted pricing_lost (reason: parser_miss, rawId: gpt-live-1) on every poll since 00:15 UTC on 2026-11-20: at 00:15, 00:30 and 00:45. The stored card is kept (keepPrior).
https://developers.openai.com/api/docs/models/gpt-live-1.md now prices two per-minute rates:
| Metric | Price | Unit |
|---|
| Active | $0.05 | minute |
| Standby | $0.02 | minute |
The page explains: "The lower standby rate applies to WebRTC sessions while inference is paused." It also says "Session duration is not rounded up to the next whole minute." The pricing page (/api/docs/pricing.md, GPT-Live section) has the same Active and Standby columns. A new gpt-live-1.1 row there has the same rates and its own model page.
parseUnitSection (src/server/providers/openai-model-docs.ts:174-215) accepts only one unsized rate per section (if (flat !== null) return null). The second flat row makes the whole model refuse, so the parse returns null.
Why it matters#
- The stored card for
gpt-live-1 bills only active minutes. The page now states a second billed quantity, so a standby-heavy session is mispriced.
gpt-live-1.1 gets no card at all.
Fix#
- Price both rates the page states as two usage quantities: active session duration and standby session duration, each billed per second, at $0.05/60 and $0.02/60. Use whatever lever names fit the rate-card language. Don't merge the two into one rate, and don't drop standby.
- A per-unit section with several unsized rows whose metric names this parser does not know should still refuse. Only
Active and Standby are stated levers.
gpt-live-1.1 should then get the same card from its own page.
- No fallback: if the page stops stating either rate, refuse as today.
- Add fixture tests from both live model pages.