Only measured availability is published.

35 live · 29 temporarily unavailable

Generated from the same public registry used by the API catalog. Last observed catalog check: 2026-10-31T22:43:50.315Z. This is a point-in-time compatibility snapshot, not a historical uptime percentage.

What the health status means

The score is a compact operational signal, not a quality benchmark. It combines success, latency, streaming completion, rate-limit responses and synthetic checks.

Publication and refresh schedule

1. Lightweight probes run every five minutes for protocol and endpoint health.
2. A round-robin model sweep runs every eight hours; each result has a visible checked timestamp.
3. A model stays public only while its check is within the freshness window, price is active and margin is at least 50%.
4. After a failure, the catalog marks the exact model unavailable; it does not reroute the request to another model.
5. Recovery requires another successful check and an audited publication action.

Privacy of measurement

Probes use synthetic prompts and never copy customer requests into shadow traffic. We retain status, latency, token counts and safe error classes for operations. Prompts, source code and response bodies are not written to logs, traces, analytics or Sentry.

For current incidents and component availability, see status.model.sale. For price calculation, see the pricing methodology.