Feature · reliability

AI API uptime SLA: what 99.9% actually means

99.9% sounds absolute. It is a monthly arithmetic claim with a window, exclusions, a credit schedule and a claims process. Here is how to read one before you sign.

Short answer. 99.9% monthly uptime allows about 43 minutes 12 seconds of unavailability in a 30-day month. The number only means something when you know the window, what counts as down, what is excluded, how the credit is calculated and how long you have to claim it. The current Plugsky Terms set Starter 99.0%, Builder 99.5%, Growth and Scale 99.9%, and Enterprise custom typically 99.95%. The pricing page headlines 99.9%; the Terms and the executed order form control. Verify the current terms.

What 99.9% means in minutes

Uptime is a percentage of time, so the allowance scales with the month: (1 − target) × minutes in the window. A 30-day month has 43,200 minutes, so 99.9% allows 43.2 minutes. A 31-day month allows 44 minutes 38 seconds; February allows less. That is why the window matters as much as the headline.

Monthly targetAllowed downtime, 30 daysWhat it feels like
99.0%7 h 12 minA working day lost across the month
99.5%3 h 36 minOne long incident, or several short ones
99.9%43 min 12 sOne bad hour, with no room for a second
99.95%21 min 36 sA short queue or a slow failover
99.99%4 min 19 sAutomatic failover territory

The percentage is not a promise that every request succeeds. A service can be “up” while your requests fail because of a bad deploy, a quota limit or an upstream model provider. Those distinctions live in the definitions.

The measurement window and what counts as down

Ask five questions. Window: calendar month, billing cycle or rolling 30 days? Metric: health endpoint, synthetic probe, 5xx rate or successful completions? Location: edge, one region or every region? Scope: account, workspace, model or endpoint? Threshold: one failed request, a 1% error rate for five minutes, or p95 latency? A useful SLA names the metric, aggregation and data source. “Monthly Uptime Percentage” is not enough. Instrument your own success rate, p50/p95 time to first token and error budget; compare them with the vendor's status page.

Exclusions, credits and the claim process

The current Terms exclude scheduled maintenance, force majeure, customer-side issues such as DNS, network or an expired card, and abuse-related suspensions. They do not explicitly address upstream model-provider outages. Because Plugsky routes to underlying providers, ask how an upstream outage is treated and get the answer in writing. Ask the same about control-plane incidents and beta features.

PlanUptime targetCredit if the target is missed
FreeBest effortNone
Starter99.0%10% of the monthly fee
Builder99.5%15% of the monthly fee
Growth99.9%25% of the monthly fee
Scale99.9%30% of the monthly fee
EnterpriseCustom, typically 99.95%Negotiated

Growth appears in the current Terms; confirm its availability and price on the pricing page. The credit is a percentage of the monthly fee, not a refund of lost revenue. Credits are the sole remedy, and a claim must go to sla@plugsky.com within 30 days with dates, times, affected resources and request logs. Put that evidence in your incident runbook before you need it.

Status history and incident RCAs

The status page shows live health, a 90-day uptime view and an incident list; the feed is at /api/status/incidents. On 20 November 2026 it listed two resolved minor incidents:

RefDateWhat happenedDuration
ST-10115 Jun 2026p95 latency on plugsky-pro reached 2.4 s after upstream throttling; traffic moved to plugsky-frontier~25 min
ST-10025 May 2026Some dashboard pages took 5–8 s after a cache eviction; no data loss12 min

That is a snapshot, not a complete archive. The metrics endpoint returned empty series on the same date, so this page does not publish a derived 90-day percentage. Ask for the raw probe data and reconcile it yourself. An incident summary is not a root-cause analysis: ask for the last two RCAs under NDA, and agree whether P1/P2 incidents get a written RCA, how quickly, and whether it covers detection, mitigation, impact and preventive actions.

Multi-region behavior

A second region is not automatically a second nine. The status page reports UAE-1, and Enterprise offers region-locked data planes, VPC, on-premises and air-gapped deployment. Ask whether the SLA is per region, per account or aggregate; whether a control-plane outage counts; whether failover is automatic; and what RTO and RPO to expect. Test it: disable the primary route, measure detection, cutover, cold start and return. A warm standby that has never served production is a hypothesis.

How to verify the SLA before signing

  • Definition: what is “unavailable”, which models and endpoints are covered, and which metric is used.
  • Window and scope: calendar month or rolling 30 days; per account, workspace, region or model; UTC boundaries.
  • Exclusions: maintenance, force majeure, customer-side issues, abuse, beta features and upstream providers.
  • Credit: percentage, cap, sole-remedy clause, cash or account credit, and how it is applied.
  • Claim: deadline, required evidence, reviewer and payment timing.
  • Operations: status page, incident feed, RCA commitment, escalation and named contacts.
  • Exit: termination assistance, data export and what happens to credits if you leave.

Then run a 30-day pilot on the topology you intend to buy. Compare your success rate and latency with the status page, file one test ticket at each severity, and ask for a sample credit calculation. An SLA you have not exercised is a document, not a control. For the commercial context, see the Enterprise overview and the current Terms.

FAQ

Does 99.9% mean 43 minutes of downtime?

In a 30-day month, 0.1% is 43 minutes 12 seconds. A 31-day month is 44 minutes 38 seconds. The SLA may define downtime differently, so use the contract definition, not the arithmetic alone.

What does an SLA credit cover?

A credit is usually a percentage of the monthly fee, not a refund of lost revenue. The current Terms set 10% for Starter, 15% for Builder, 25% for Growth, 30% for Scale and negotiated Enterprise credits; credits are the sole remedy.

Are scheduled maintenance and upstream outages excluded?

Scheduled maintenance, force majeure, customer-side issues and abuse suspensions are explicitly excluded. The Terms do not explicitly address upstream model-provider outages; ask how they are treated and get the answer in writing.

Where is the status history?

The status page shows live health and an incident list; the incident JSON is at /api/status/incidents. When checked on 20 November 2026 it listed two resolved minor incidents, not a full historical archive.

Does a second region double the SLA?

Not automatically. The SLA may be measured per region, per account or in aggregate, and failover may be manual. Verify the aggregation rule and test failover before you count on it.

Sources

Primary sources checked 20 November 2026. Provider terms, limits and prices change; verify the current version before you commit.

Reviewed 20 November 2026. The SLA schedule is quoted from the current Terms; the pricing page headline and the plan-level Terms differ. Confirm the current version and your order form.