Azure OpenAI
As of Jun 9, 22:00 UTC · revalidates every 60 seconds · pipeline ingests every 5 minutes.
Impact-weighted uptime
Azure OpenAI published 2.7 incidents per 30 days on average to its status feed across the last 6 months (n=16). We surface this number instead of classifying providers because uptime feeds vary in publishing volume for reasons we cannot judge from outside; compare the uptime numbers below across providers with this denominator in mind.
Uptime weighted by per-incident UIS over the window, computed from incidents this provider published to its own feed. n splits into live (observed in real time) and historical (re-derived from the provider's own archive). Not directly comparable across providers — see the feed-volume disclosure above. How this is computed.
Multi-region reachability + latency
Probes hit the provider's status page from each region every 15 minutes. Per-region latency variance is a signal worth watching even when the provider's own status feed reports operational. When ≥1 region reports unreachable while ≥1 other reports reachable, the disagreement also feeds the §6.1 multi-source confirmation gate.
Published SLA compliance
Published SLA: 99.90% · Azure OpenAI Service, monthly · source (www.microsoft.com) →
Computed by comparing our impact-weighted uptime against the provider's own published target. We are not making a cross-provider claim here — each provider is held to its own commitment. Measurement methodology is the same impact-weighted formula used elsewhere on this page.
Component classification · 4 rules
Components classified by class (primary inference, secondary API, auxiliary). Flagship models tracked: 3.
Latency probe · parallel-run, methodology v1.0
Alt-signal observations · last 24 hours
Unconfirmed signal observations from the last 1440 minutes. None are published incidents; they corroborate confirmed incidents and feed the multi-source gate. Single-source signal is intentionally not auto-posted.
Per-model community reports · last 24 hours
Per-model claims are gated: published only when the provider names the model OR at least two GitHub SDK issues name it with an HTTP 5xx pattern within 30 minutes. Other mentions are surfaced as community reports without a per-model claim.