Managed Inference Providers: Uptime SLAs and Credits
The short answer
If a contractual uptime SLA with service credits is the requirement, buy inference through a hyperscaler or a reserved-capacity tier — not pay-as-you-go APIs. As of October 19, 2026: Amazon Bedrock publishes a 99.9% monthly uptime SLA per region, Azure OpenAI 99.9%, and Google’s Gemini online inference on Vertex AI 99.5%. OpenAI gives 99.9% only on Scale Tier; Together AI publishes 99% on Provisioned Throughput. Fireworks serverless and Anthropic’s standard tier have no SLA; Fireworks, Baseten and Anthropic negotiate SLAs in enterprise contracts.
The comparison
| Provider / tier | Published uptime commitment | Service credits | How you qualify |
|---|---|---|---|
| Amazon Bedrock | 99.9% monthly, per AWS region | 10% (99.0–<99.9%) · 25% (95.0–<99.0%) · 100% (<95.0%) of that region’s Bedrock charges | Any account; file “SLA Credit Request” with request logs by end of the second billing cycle |
| Gemini on Vertex AI (Gemini Enterprise Agent Platform online inference) | 99.5% for generateContent / streamGenerateContent; 95% for models designated shorter-availability | 10% (99.0–<99.5%) · 25% (95.0–<99.0%) · 50% (<95.0%) | Any Google Cloud project; Provisioned Throughput adds a latency target |
| Azure OpenAI (Foundry Models) | 99.9% | Per Microsoft’s Online Services SLA | Pay-as-you-go or provisioned throughput units (PTUs) |
| OpenAI API — Scale Tier | 99.9% uptime, plus a per-model throughput target (e.g. 99% of requests above 50 tokens/s) | Contract terms | Prepaid units of tokens per minute, billed per day |
| OpenAI API — standard | None | — | — |
| Anthropic API | Standard tier best-effort; Priority Tier no longer sold | Contract terms | Contact sales for guaranteed capacity |
| Together AI — Provisioned Throughput | 99% uptime SLA | Contract terms | Committed capacity, token-based pricing |
| Fireworks | Serverless: no SLA; enterprise reserved GPUs: SLA by contract | Contract terms | On-demand or reserved deployment; P0 support response within 1 hour |
| Baseten | Custom SLAs on Enterprise | Contract terms | Enterprise plan |
What the numbers mean in practice
99.9% allows about 43 minutes of downtime in a 30-day month; 99.5% allows about 3.6 hours; 99% about 7.2 hours. The gap between Bedrock and Vertex AI on the headline number is real, but read the definitions before comparing:
- How downtime is counted. Google counts a minute as down only when more than 5% of valid requests return HTTP 5xx, and only in runs of five or more consecutive minutes. A burst of 429 rate-limit errors is not downtime on any of these SLAs.
- What the credit is worth. Credits are a share of what you paid for that service in that region or project that month, applied to future bills. AWS issues nothing under $1 and states credits are the “sole and exclusive remedy.” None of them covers the revenue your product lost.
- You must claim. AWS wants the words “SLA Credit Request”, the region, the affected 5-minute intervals and your own request logs, by the end of the second billing cycle after the incident. Keep per-request logs with timestamps and status codes, or you cannot claim.
- Capacity is a separate promise. An uptime SLA does not guarantee you get throughput when the region is busy. That is what OpenAI Scale Tier units, Azure PTUs, Vertex Provisioned Throughput and Together Provisioned Throughput sell — and why the throughput targets sit next to the uptime number.
Picks by situation
- You need a signed SLA without negotiating: Amazon Bedrock (99.9%, the most generous credit at the bottom of the scale) or Azure OpenAI (99.9%). Both serve Anthropic or OpenAI models with the cloud’s SLA attached.
- Gemini is your model: Vertex AI at 99.5%, plus Provisioned Throughput if latency matters.
- OpenAI models with guaranteed speed: Scale Tier, or Azure PTUs if you are already on Azure.
- Open-weight models with an SLA: Together Provisioned Throughput (99% published) or a Baseten / Fireworks enterprise contract; compare dedicated GPU prices in dedicated inference endpoint providers.
- Prototype or internal tool: pay-as-you-go APIs are fine; add a fallback route instead of an SLA.
The part credits do not fix
The practical answer to an outage is two providers behind a gateway with automatic failover, not a bigger credit. Route the same model through two hosts where possible (Claude on the Anthropic API and Bedrock; an open model on two of Together, Fireworks and Baseten), and keep a cheaper fallback model warm. See best LLM gateways ranked for the routing layer and low-latency inference providers if speed, not uptime, is the real requirement. Price the models on current API prices.
Last verified: October 19, 2026. Uptime and credit terms are from the vendors’ SLA and documentation pages; contract SLAs vary by customer.