We poll providers about every 5 minutes and score models for consumers — not for vanity. Health dots are “right now.” Availability bars summarize ~15‑minute windows with honest thresholds (green requires ≥90% healthy polls; blue needs multi-provider HA in a reliable window).
Independent providers run inference. This page reports whether models are routable right now and how stable that has been over the past 14 days. We do not invent live bid prices or model counts — we measure provider self-reports against on-chain bids.
models[] health self-report (we never force a re-probe).One color per model from its current bids:
| Dot | Status | Rule |
|---|---|---|
| Blue | redundant | ≥2 healthy bids (multi-provider HA) |
| Green | operational | Exactly 1 healthy bid |
| Yellow | degraded | No healthy; ≥1 degraded (HTTP 429 / rate-limited) — still routable |
| Slate | unknown | Reachable but legacy (< v7.5) or no usable self-report |
| Hollow | offline | Only hard failures / unreachable |
Routable marketplace lists include only blue / green / yellow models.
Each bar is a ~15‑minute window over several 5‑minute polls. We compute the share of polls
that were redundant (r), operational (h), degraded (d),
unknown (u), or offline (o), then pick the first matching color:
| Color | Condition | What it means |
|---|---|---|
| Blue | ≥50% redundant and ≥90% healthy | Multi-provider HA was the majority of a reliable window |
| Green | ≥90% healthy (redundant + operational) | Reliably available (single provider is OK) |
| Yellow | ≥50% routable (incl. 429) | Soft capacity / intermittent health |
| Slate | ≥50% unknown | Reachable but we can’t trust health yet |
| Orange | ≥25% routable | Flaky — up some of the time |
| Red | Otherwise | Mostly failed |
| Empty | No data, pre-birth, or 100% offline | Nothing to score |
With ~3 polls per bar, ≥90% healthy means essentially all polls in the window must be good — one miss drops green to yellow. That is intentional: we would rather look strict than paint a flaky window green.
Overall color is health-weighted — redundant counts as healthy, not a separate “less good” bucket. Marketplace availability changes (models joining or leaving the routable set) are not shown as an “open incident” alarm — subscribe to the network feed (high-signal) or verbose feed for those. A notice appears only when status is major.
| Banner | Rule |
|---|---|
| Green | ≥90% of scored models healthy (redundant + operational) |
| Orange | ≥50% and <90% healthy |
| Red | <50% of scored models healthy |
Scored = redundant + operational + degraded + offline. Unknown/legacy is excluded from the denominator (reachable-but-untrusted is not an outage). A few soft-degraded (429) models do not force the banner orange when ≥90% remain healthy.
Separate from bar color: share of sweeps where the model was routable over 14 days. Badge tint: green ≥99.5%, yellow ≥95%, orange ≥34%, else red.
Hover any availability bar for the poll mix in that window (e.g. 2/3 redundant · 1/3 operational).