Search “LLM API SLA” and you get tables that disagree with each other. This post skips the tables and reads the source: each provider’s own SLA or terms page, fetched on 15 November 2026. For each one we note the uptime target, what counts as downtime, the service credit, any latency target and which tier is covered. BroadRouter is BroadNet’s website, so BroadNet is in the list, and it is one of the providers without an SLA on self-serve.
What does an LLM API SLA actually promise?
An SLA is a written target with a remedy. It does not make the API stay up. It says what happens on your bill when it doesn’t. Every SLA below has the same four parts:
- A target. Usually a monthly uptime percentage per region, project or resource.
- A definition of downtime. Almost always server errors only: 500 or 5XX responses to valid requests. A 429 rate limit is not downtime. Neither is a slow answer or a wrong answer.
- A remedy. A percentage of that month’s bill for the affected service, credited to future bills and capped.
- A claim process. You ask for the credit, with your own logs, inside a deadline.
The percentages translate into minutes like this, for a 30-day month of 43,200 minutes:
| Monthly uptime | Downtime allowed |
|---|---|
| 99.9% | about 43 minutes |
| 99.5% | 216 minutes |
| 99.0% | 432 minutes |
| 95.0% | 36 hours |
Which LLM APIs publish an SLA?
All rows read from the provider’s own page on 15 November 2026. “Not found” means the page we checked states no SLA; it does not rule out a private contract.
| Provider and service | Uptime target | Counts as downtime | Credits | Latency SLA | Who is covered |
|---|---|---|---|---|---|
| Microsoft Azure, Foundry Models (includes Azure OpenAI models) | 99.9% | Requests returning an error code, averaged per minute | 10% below 99.9%, 25% below 99% | Yes, for Provisioned Managed deployments and Priority Processing | Foundry Models sold by Azure; Developer tier excluded |
| Amazon Bedrock | 99.9% (credits start below it) | Requests that return a 500 error code, per 5-minute interval, per region | 10%, 25% below 99.0%, 100% below 95.0% | None in the SLA | Bedrock in each AWS region |
| Google, Gemini online inference API | 99.5% (95% for models with shorter availability) | Over 5% 5XX error rate for 5 or more consecutive minutes | 10%, 25% below 99.0%, 50% below 95.0% | Yes, under Provisioned Throughput on the global endpoint | generateContent and streamGenerateContent |
| OpenAI, Scale Tier | 99.9% | Not defined on the public page | Not stated on the public page | 99% of time above a per-model tokens-per-second floor | Enterprise customers, token units bought for at least 30 days |
| OpenAI, standard pay-as-you-go | Not found | – | – | – | – |
| Anthropic API | Not found; terms say “as is” and “as available” | – | – | – | – |
| Fireworks AI, Reserved Throughput | 99.9% uptime terms | Not stated on the pricing page | Not stated | Throughput SLA | Reserved capacity only; overage not covered |
| Together AI | Serverless: none stated. “Provisioned Throughput: Token-based capacity with SLAs” (terms not published on the page) | – | – | – | Ask sales for the SLA terms |
| BroadNet Inference, self-serve | None: “no availability commitment” | – | – | – | Service levels only in a separate written agreement |
How do the hyperscaler SLAs differ?
Microsoft Azure. Microsoft’s Learn page says Foundry Models sold by Azure “include all Azure OpenAI models” and are “covered by Azure service-level agreements”. The consolidated SLA (October 2026 edition) defines Error Rate per minute and Uptime Percentage as 100% minus the average error rate. Credits are 10% below 99.9% and 25% below 99%. Credits apply only to Foundry Models sold by Azure, and “The Developer tier of the Foundry Models service is not covered by this SLA.” For Provisioned Managed deployments and Priority Processing, a second SLA measures tokens per second against a per-model target in the product docs, with a 10% credit if attainment falls below 99%.
Amazon Bedrock. AWS commits to “commercially reasonable efforts” to meet the uptime table. Availability is “the percentage of Requests processed by Amazon Bedrock that do not fail with Errors”, and an Error is “any Request that returns a 500 error code”. It is calculated per region, and intervals with no requests count as fully available. Exclusions include customer actions such as “scaling of provisioned throughput” and model software causing “repeated model crashes or an inoperable model”. The SLA has no latency target.
Google Cloud. The Gemini online inference SLA is the strictest about what counts. Downtime needs more than a 5% error rate of 5XX responses, lasting “five or more consecutive minutes”. Credits are capped at 50% of the month’s bill for the service, you must notify support within 30 days with log files, and the SLA is your “sole and exclusive remedy”. Grounding with Google Search, quota errors and deadlines shorter than the server default are excluded. Under Provisioned Throughput, a latency SLA covers some Gemini 2.5 models on the global endpoint, measured as p50 tokens per second per 5-minute interval. If you deploy your own open model on Vertex AI instead, the separate Vertex AI Platform SLA covers custom model online prediction at 99.5%, but only “for models deployed on 2 or more nodes”.
What about OpenAI and Anthropic directly?
OpenAI. The Scale Tier page says: “Scale Tier traffic offers a 99.9% uptime SLA and prioritized compute.” Its table pairs that with a latency SLA per model, such as “99% > 50 tokens per second”, calculated as p50 request latency per 5 minutes. The page says “This offering is available to Enterprise customers”, that each token unit is bought “for a minimum of 30 days”, and that Scale Tier covers models released before GPT-5.6, with Reserved Tier for later models. We found no uptime SLA attached to standard pay-as-you-go use. (openai.com blocked our fetcher, so we read the page from the Internet Archive’s 7 November 2026 capture.)
Anthropic. The public Commercial Terms (effective 17 June 2025) contain no SLA or service credits. Section L.2 provides the services “AS IS” and “AS AVAILABLE” and disclaims that use “WILL BE UNINTERRUPTED”. Enterprise agreements may differ, but they are not public.
What does an SLA not buy you?
- Uptime itself. A credit is paid after the outage. Your users still saw it.
- Capacity. Throttling (429) is outside every uptime definition above. Google explicitly excludes “quotas applied by the system”. See how to handle 429 errors.
- Speed. Only the provisioned or reserved tiers carry a tokens-per-second target. On standard traffic, a slow answer is a successful answer. Latency metrics explained.
- Quality. A 200 response with a wrong answer is not downtime anywhere.
- Much money. Credits are a share of one service’s monthly bill, often capped, and you must claim them with your own logs.
How do you build on an API with no SLA?
Most teams build on pay-as-you-go APIs with no SLA. That works if the application expects failure:
- Set timeouts. Stream long answers and set a client timeout per request, so a hung call fails fast. Streaming guide.
- Retry with backoff and jitter on 429 and 5XX, and cap retries. Errors and retries.
- Keep a second provider configured. With OpenAI-compatible APIs, a fallback is a second base URL, key and model ID. How to fail over between LLM APIs.
- Log request IDs, timestamps and status codes. You need them for any SLA claim, and to see your real error rate.
- Buy a contract where you need one. If uptime is a hard requirement, use a provider tier with a written SLA, or a private deployment with agreed service levels. What to plan before a private endpoint.
Where BroadNet Inference fits
BroadRouter is BroadNet’s own website. BroadNet’s self-serve API, pay-as-you-go and plans alike, has no SLA. Its Terms of Service (§7.4) say: “We aim to keep the API available and publish what we measure, but self-serve has no availability commitment.” And: “Service levels exist only in a separate written agreement.” On this point Azure, AWS and Google offer more: a written uptime target with credits. A pay-as-you-go BroadNet key allows 30 requests a minute and 4 in flight. BroadNet’s own description of its serving setup is in BroadNet’s GPU hardware and serving approach, which we report as BroadNet’s account, not as a commitment.
If your workload can run with retries and a fallback, BroadNet’s OpenAI-compatible endpoint is one more base URL to configure: https://inference.broadnet.ai/v1. If you need contracted service levels, ask BroadNet about a written agreement before you rely on one. Read the API reference or create an account.
