Availability and Service Levels
How availability is measured, when failed requests are billed, and what an enterprise agreement can cover.
This page explains how the platform measures availability, how failed requests are billed, and what an enterprise agreement can add.
Self-serve availability
Model availability naturally follows the upstream model provider: when a provider degrades, requests routed to it can fail, and no gateway can prevent that.
What the platform does control:
- Provider failover. Requests are scheduled across upstream providers so that a single provider's degradation has a limited effect on your traffic.
- No charge without usage. A request that produces no billable usage is never charged, regardless of whether the failure came from the platform or from an upstream provider.
- Visible measurement. Each model page shows the model's observed availability over the last 24 hours, computed with the definition below, so you can judge reliability from data rather than from a claim.
For more detail on how availability and reliability are handled, contact your dedicated account manager, or email support@modelrover.ai.
How availability is measured
The availability figure on a model page is measured as follows.
| Item | Definition |
|---|---|
| Window | Rolling 24 hours, displayed in 15-minute buckets |
| Counted as success | Completed requests, and requests cancelled by the caller |
| Counted as platform fault | Failed requests whose error is not caused by the caller's request or credentials |
| Excluded from the calculation | Caller-caused outcomes: invalid, disabled, or expired API key, insufficient balance, rate or concurrency limits, unsupported model, capability or parameter, invalid request, oversized request, content policy rejection, missing or unusable assets, unknown task |
| Excluded from the calculation | Requests truncated by a balance limit |
| Status thresholds | Operational at 95% or above, degraded between 80% and 95%, outage below 80% |
Because caller-caused outcomes are excluded rather than counted as failures, the figure reflects platform and upstream reliability for valid requests. If you see a drop against a specific request ID, check request logs for the stable error code before drawing conclusions about availability.
Billing for failed requests
Billing depends on whether billable usage was produced, not on whether the request succeeded.
- No billable usage: not charged. This covers rejected requests, exhausted capacity, authentication and balance failures, and upstream errors that returned no usable output.
- Usage already produced: charged at actual usage, even if the stream was interrupted afterwards.
- Upstream cost of a failed attempt: when a request fails without delivering usable output, any cost the upstream provider charges for that attempt is borne by the platform, not passed to you.
Retries that produced billable output contribute to the final usage of the request. The request log is the source of truth for the settled amount of any call.
See Pricing and Billing for how charges are calculated.
Rate limits
Rate limits apply at several layers at once. Every layer must pass, and the strictest value for a dimension decides — reaching a limit at any layer returns 429 immediately, before the request is routed.
| Layer | Dimensions | Configured by |
|---|---|---|
| Platform default | Requests per minute, concurrency | Platform operator |
| Account | Requests per minute, concurrency | Platform operator, shared by every API key on the account |
| API key | Requests per minute, tokens per minute, concurrency | You, in the console |
| Per model, per account | Requests per minute, concurrency | Platform operator, shown on each model page |
Account and API key limits are shared, not additive. An account has one concurrency and requests-per-minute budget. Keys configured with their own limits draw from that same budget: ten keys each configured for 2000 requests per minute do not add up to 20000 — the account limit caps the total.
Per-model limits are counted for each account separately, so one account reaching its quota on a model does not reduce another account's quota.
Requests per day is not currently enforced and is shown as unlimited. Exceeding a limit returns 429 with a Retry-After header; when every upstream channel for a model is saturated at the same moment, the request instead returns 503 with the channel_capacity_exhausted error code.
Rate-limit and capacity rejections produce no billable usage and are excluded from the availability calculation above.
Enterprise service levels
An enterprise agreement can add commitments that the self-serve platform does not provide. The typical scope:
- An availability target for the platform, measured with the definition above
- Service credits when the target is missed, including the trigger threshold and the claim window
- Support response targets by severity, with an escalation path
- Contractual terms your procurement and security review require, such as a data processing agreement
Service levels are negotiated per agreement. To request one, email business@modelrover.ai with your expected volume, the models you depend on, and the commitments your review requires. We reply with a draft agreement rather than a published template, because the terms are set per customer.
Related
- Errors, Limits, and Retries for the status codes and error codes behind these outcomes
- Pricing and Billing for how charges are calculated
- FAQ and Support for support channels
Help us improve this page
Found something unclear, outdated, or incorrect?
Last updated on