Short answer. Fifteen providers publish a status page. Thirteen expose a machine-readable feed you can poll; two publish only a service-health dashboard. This tracker publishes the endpoints and starts its own 90-day uptime series on 20 November 2026. It does not backfill history, because a status page archives incidents, not a continuous uptime measurement.
Baseline: 20 November 2026. The uptime column is empty by design. The first complete 90-day window ends on 18 February 2027. Until then, the honest number is “not yet measured”, not a reconstructed percentage.
What this tracker measures
Uptime is usually reported as a percentage with no definition attached. Here it means the share of one-minute probes that received a successful HTTP response from the provider’s inference endpoint, measured from one region, excluding probes we failed ourselves. That is a narrower claim than “the service was up”: a provider can be up for chat and down for a specific model, or up in one region and down in another.
The tracker therefore records four things separately: the status page state, the probe result, the incident feed, and the model and region affected. A single percentage collapses all four and hides the only part an engineer can act on.
The 15 providers
| Provider | Status page | Machine-readable | History on page |
|---|---|---|---|
| OpenAI | status.openai.com | Statuspage JSON and RSS | Incident archive |
| Anthropic | status.anthropic.com | Statuspage JSON and RSS | Incident archive |
| Google Cloud (Vertex AI) | status.cloud.google.com | Incident JSON | Incident archive |
| Microsoft Azure (Azure OpenAI) | azure.status.microsoft | Service health feed | Incident archive |
| Amazon Web Services (Bedrock) | health.aws.amazon.com | Health dashboard | Incident archive |
| Mistral | status.mistral.ai | Statuspage JSON and RSS | Incident archive |
| Cohere | status.cohere.com | Statuspage JSON and RSS | Incident archive |
| Groq | groqstatus.com | Statuspage JSON and RSS | Incident archive |
| Together AI | status.together.ai | Statuspage JSON and RSS | Incident archive |
| Fireworks AI | status.fireworks.ai | Statuspage JSON and RSS | Incident archive |
| OpenRouter | status.openrouter.ai | Statuspage JSON and RSS | Incident archive |
| Replicate | status.replicate.com | Statuspage JSON and RSS | Incident archive |
| Hugging Face | status.huggingface.co | Statuspage JSON and RSS | Incident archive |
| xAI | status.x.ai | Statuspage JSON and RSS | Incident archive |
| Perplexity | status.perplexity.ai | Statuspage JSON and RSS | Incident archive |
Checked 20 November 2026. A status page can move off Statuspage or change its feed without notice; the endpoint, not the label, is what you should pin. Providers that publish only a dashboard are marked as such.
Why the 90-day number starts today
There are three ways to produce a 90-day uptime figure, and only one of them is honest. You can measure it yourself, you can read it from a provider that publishes it, or you can infer it from incident posts. The third method produces a number that looks precise and is not: incident posts give durations with minute-level rounding, they often cover a component rather than the API, and a provider may post one incident for an outage that hit several models.
So the series starts now, with a defined probe, a fixed region and a published method. The table below is the publication schedule, not a result.
| Milestone | Date | What is published |
|---|---|---|
| Baseline | 2026-11-20 | Probe definition, provider list, incident taxonomy |
| Day 30 | 2026-12-20 | Probe count, error classes, partial-window uptime |
| Day 60 | 2027-01-19 | Second partial window, first incident comparison |
| Day 90 | 2027-02-18 | First full 90-day uptime, per provider and per model |
If a provider publishes its own 90-day figure before then, we will show it beside ours and label the source. Two numbers with different methods are more useful than one number with no method.
Polling a status feed
Statuspage exposes a summary endpoint that returns the current indicator and the active incidents. It is a cheap health check and a good fallback when your own probe fails: if the provider says “major outage” and your probe agrees, you can stop debugging your client.
import json, urllib.request
def status(host):
url = f"https://{host}/api/v2/status.json"
with urllib.request.urlopen(url, timeout=5) as r:
d = json.load(r)
return d["status"]["indicator"], d["status"]["description"]
for host in ("status.openai.com", "status.anthropic.com"):
print(host, status(host))
Two cautions. A status API is not an SLA: it can lag the real incident by minutes, and it reflects the provider’s own view. And a 200 from the status page does not mean the inference endpoint is healthy. Probe the endpoint you actually call, with the model you actually use.
Incident taxonomy
| Status page state | What it usually means | What to record |
|---|---|---|
| Operational | No known issue | Probe success, latency |
| Degraded performance | Elevated latency or error rate | p95 latency, 5xx rate, model |
| Partial outage | Some models or regions failing | Model ID, region, error code |
| Major outage | Broad failure | Start, end, affected endpoints |
| Maintenance | Planned unavailability | Window, notice period |
Map the provider’s wording onto these five states before you compute anything. A “degraded performance” incident that lasts four hours is not the same as four hours of downtime, and counting it as downtime will make every provider look worse than it is.
Methodology
- Probe. One request per minute per provider, from one region, with a 10-second timeout. Success means HTTP 200 and a valid JSON body, not just a TCP connection.
- Model scope. One small text model per provider, pinned at the start of the window. A model retirement restarts that provider’s series and the change is logged.
- Exclusions. Probes that fail because of our own network, DNS or credentials are excluded and counted separately.
- Reporting. Uptime is successful probes divided by eligible probes. Latency is reported separately as p50 and p95; a fast failure is still a failure.
The method has limits. One region is not global availability, one model is not the whole catalogue, and a minute-level probe can miss a 20-second outage. The point is consistency: the same probe, the same model, the same arithmetic, published with the raw counts so you can recompute it.
Sources
Links checked 20 November 2026. We cite the primary page for each figure. Where a provider publishes a price behind a login, we say so and do not guess.
- OpenAI status — current state, incident history and Statuspage feed
- Anthropic status — current state and incident history
- Google Cloud status — incident JSON and per-product history
- AWS Health Dashboard — service health and event history
FAQ
Where can I see 90-day uptime for AI APIs?
This tracker publishes its first full 90-day window on 18 February 2027, measured from 20 November 2026. Before that date we show partial windows and the raw probe counts. Provider status pages archive incidents but do not publish a continuous uptime series.
Do all AI providers have a status API?
No. Twelve of the fifteen providers here expose a Statuspage-style JSON or RSS feed. Google Cloud publishes an incident JSON, while AWS and Azure expose service health dashboards that are harder to poll from a script.
Does a green status page mean the API is working?
No. Status pages report the provider’s view and often lag the incident. A model can fail while the page is green, especially during a partial outage or a regional issue. Probe the endpoint and model you actually call.
How do you count a degraded-performance incident?
We record it as degraded, not as downtime. Uptime counts failed probes only; elevated latency is reported separately as a p95 series. Collapsing the two makes comparisons between providers meaningless.
Why not reconstruct 90 days of history from incident posts?
Because incident posts are rounded, often component-specific and not a continuous measurement. Reconstructing a percentage from them would look precise without being comparable. We start the series at the baseline and publish the method instead.
Use the data, then test it on your own workload
Plugsky publishes its plan caps and model limits. Start free with no card, or compare the flat plans against per-token billing.
Start Free → See plans