AI API uptime, measured independently
Free API key (higher limits) →
Status pages are written by the provider; this one is measured. Every 5 minutes InfraRabbit sends one unauthenticated request to 24 public AI API hosts and records whether, and how fast, they answer. Newest measurement: (UTC), from one vantage (werk-nl); first measurement 2026-11-18T00:57:24+00:00.
| Host | Now | 24h uptime | 7d uptime | 30d uptime | 24h latency p50 / p95 | Incidents, 30d |
|---|---|---|---|---|---|---|
| OpenAI api.openai.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1018 / 1589 ms | 0 |
| Anthropic api.anthropic.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
891 / 1385 ms | 0 |
| Google Gemini API generativelanguage.googleapis.com |
up HTTP 403 | 99.65% 288 windows, 1 down | collecting 306 windows, 15% of the period covered, 1 down | collecting 306 windows, 4% of the period covered, 1 down |
822 / 1377 ms | 0 |
| Mistral api.mistral.ai |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
870 / 1377 ms | 0 |
| xAI api.x.ai |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1003 / 1485 ms | 0 |
| DeepSeek api.deepseek.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
990 / 1476 ms | 0 |
| Groq api.groq.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
995 / 1485 ms | 0 |
| Together AI api.together.xyz |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
987 / 1415 ms | 0 |
| Fireworks AI api.fireworks.ai |
up HTTP 401 | 99.31% 288 windows, 2 down | collecting 306 windows, 15% of the period covered, 2 down | collecting 306 windows, 4% of the period covered, 2 down |
805 / 1289 ms | 0 |
| OpenRouter openrouter.ai |
up HTTP 200 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
785 / 1311 ms | 0 |
| DeepInfra api.deepinfra.com |
up HTTP 200 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1504 / 2008 ms | 0 |
| Cerebras api.cerebras.ai |
up HTTP 403 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
893 / 1313 ms | 0 |
| SambaNova api.sambanova.ai |
up HTTP 200 | 99.65% 288 windows, 1 down | collecting 306 windows, 15% of the period covered, 1 down | collecting 306 windows, 4% of the period covered, 1 down |
1295 / 3694 ms | 0 |
| Perplexity api.perplexity.ai |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
903 / 1380 ms | 0 |
| Cohere api.cohere.com |
up HTTP 401 | 99.31% 288 windows, 2 down | collecting 306 windows, 15% of the period covered, 2 down | collecting 306 windows, 4% of the period covered, 2 down |
894 / 1386 ms | 0 |
| Hugging Face Inference Providers router.huggingface.co |
up HTTP 200 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
871 / 1287 ms | 0 |
| Replicate api.replicate.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
963 / 1400 ms | 0 |
| fal fal.run |
up HTTP 401 | 99.31% 288 windows, 2 down | collecting 306 windows, 15% of the period covered, 2 down | collecting 306 windows, 4% of the period covered, 2 down |
1114 / 1500 ms | 0 |
| ElevenLabs api.elevenlabs.io |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
915 / 1392 ms | 0 |
| Deepgram api.deepgram.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1221 / 1637 ms | 0 |
| AssemblyAI api.assemblyai.com |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1197 / 1758 ms | 0 |
| AWS Bedrock Runtime (us-east-1) bedrock-runtime.us-east-1.amazonaws.com |
up HTTP 404 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1006 / 1408 ms | 0 |
| Azure AI services (East US) eastus.api.cognitive.microsoft.com |
up HTTP 200 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1136 / 1589 ms | 0 |
| Lambda cloud.lambda.ai |
up HTTP 401 | 100.00% 288 windows | collecting 306 windows, 15% of the period covered | collecting 306 windows, 4% of the period covered |
1012 / 1517 ms | 0 |
How it is measured
Every 5 minutes one unauthenticated HTTPS GET per host, to a cheap documented path (usually the model list), from one vantage. We time DNS, the TCP connect, the TLS handshake and the time to the first byte of the answer, and record the HTTP status. No API key is used and no model is called, so it costs the providers and us nothing. An answer of 401, 403 or 200 (the documented reply of an unauthenticated call) inside 10 seconds is up; slower than three times the host's own 7-day p95 is degraded; a timeout, a DNS, connection or TLS error is down; a 5xx is degraded the first time and down when it repeats in the next window. A status we did not expect (a moved path) is reported as unexpected and left out of the uptime figure. If nearly every host fails to connect in the same run the run is marked probe_fault (that is our network) and left out as well. Uptime is (up + degraded) / (up + degraded + down) over the windows we actually measured; coverage shows how many of the possible windows exist. An incident is two or more consecutive down windows. The raw rows are append-only and never edited.
What this does not prove
What this proves: from this vantage (one IPv4 network, one location) the provider's DNS, TCP, TLS and edge or gateway answered, and how fast. What it does NOT prove: that a model generates, that your key, quota or region works, that other regions or IP ranges are up, or that a failure we saw is global. A provider can be fully up for this probe while inference is failing, and the reverse. Treat a down window as 'unreachable from here', and use several sources (and the provider's own status page) for a claim.
Probed hosts and paths are in the JSON. The probe identifies itself as InfraRabbitBot/1.0 (+https://infrarabbit.com/bot), sends one request per host per round and never retries. A host that wants out can write to [email protected] and is removed.
Data: /v1/uptime (24 h, 7 d, 30 d per host), /v1/uptime/<host>/history (daily rows and incidents), MCP tool get_uptime at https://infrarabbit.com/mcp. CC BY 4.0.