Azure OpenAI in Foundry Models SLAMicrosoft Azure · AI models · Microsoft Azure · AI models
Promise vs reality
Azure publishes post-incident reviews as web pages only, with no machine-readable history, so observed uptime is not computed. How this is measured
All incidents, weighted: 100% · Downtime- Promised
- 99.9%
- Observed, 365d
- n/a
- Major incident time
- n/a
- Credit on first breach
- 10%
Commitments and credits
- Foundry Models (Sold by Azure), availabilityFoundry Models Sold by Azure (including Azure OpenAI models); Developer tier not covered99.9%
Monthly uptime below, credit of the bill
- Below 99.9%10%
- Below 99%25%
- Provisioned Managed Deployments and Priority Processing, latencyLatency attainment (Tokens Per Second vs. SLA Latency Target Value) for models with a documented latency target99%
Monthly uptime below, credit of the bill
- Below 99%10%
Terms
- Measured
- Uptime Percentage per 'Applicable Period': for metered pay-as-you-go services, the 30 days prior to and including the first day of the Incident; for other services, the calendar month. Uptime = 100% minus Average Error Rate (per-minute share of requests returning an Error Code).
- Credit cap
- Maximum published credit is 25% (availability) / 10% (latency); overall cap: Service Credits in any billing month will not exceed the monthly service fees for that Service or Service Resource in the Applicable Period
- Claim window
- Within 60 days of the Incident (Azure)
- How to claim
- Submit the claim to Microsoft customer support with incident description, time and duration of downtime, affected resource names, number/location of affected users and error descriptions
Not covered
- Factors outside Microsoft's reasonable control (natural disaster, war, government action, network failures outside Microsoft data centers)
- Services, hardware or software not provided by Microsoft
- Preview, pre-release, beta or trial versions, or purchases made with subscription credits
- Customer failure to follow required/supported configurations or published guidance
- Exceeding reasonable use or quotas, or throttling of suspected abuse
- Scheduled Downtime (at least 5 days' notice) and customer-initiated restart/stop/failover/scale operations
Foundry Models ... Service Credit: Uptime Percentage Service Credit < 99.9% 10% < 99% 25% Service Level Exceptions: The Service Credits are only applicable to Foundry Models Sold by Azure. The Developer tier of the Foundry Models service is not covered by this SLA.
The September 2026 SLA document has no standalone 'Azure OpenAI' section; Azure OpenAI models are covered as Foundry Models Sold by Azure (other Foundry sections say 'A separate SLA applies to Azure OpenAI Service'). The latency tier's 'uptime' field holds the 99% latency Attainment Percentage, not an availability figure.
Evidence: incidents, last 365 days
Azure publishes post-incident reviews as web pages only, with no machine-readable history, so observed uptime is not computed.
Other ai models SLAs
ServiceDowntime vs allowed
Amazon BedrockAWS99.9% / 99.178%
Vertex AI (Gemini API on Vertex)Google Cloud99.9% / 100%
Azure OpenAI in Foundry ModelsMicrosoft Azure99.9% / n/a
Claude API Priority TierAnthropic99.5% / 98.8%
Cohere API (SaaS: Command, Embed, Rerank)Coherenone / 100%
Hugging Face Inference EndpointsHugging Facenone / n/aNo public SLA
Promised in the SLAObserved on the vendor's status pagePartial historyShortfall
Summaries of published SLAs; the contract you sign governs. Logos via logo.dev; trademarks belong to their owners.