Skip to content
QuickStatus.dev

Uptime rankings

LLM API uptime rankings

How often the major LLM APIs answered a real request this week. Every 5 minutes we ask each provider's lowest-cost model for one token, from five regions, with our own key.

Uptime this week

Week 2026-W47 so far, as of Sat 21, 07:46 UTC. Live and not final: the ranking freezes on Monday at 00:05 UTC. Updated every 5 minutes.

  1. –Coheresince Thu 19, 12:00 UTC100%
  2. –DeepSeeksince Thu 19, 12:00 UTC100%
  3. –Google Geminisince Thu 19, 12:00 UTC100%
  4. –Mistralsince Thu 19, 12:00 UTC100%
  5. –Moonshot AI (Kimi)since Thu 19, 12:00 UTC100%
  6. –OpenAIsince Thu 19, 12:00 UTC100%
  7. –xAIsince Thu 19, 12:00 UTC99.61%
  8. –AnthropicNo results yet–
  • Answered (hover a day for its uptime)
  • Confirmed outage: 2+ regions failed on 2 runs in a row
  • One run failed (shown, not counted)
  • Not run yet

Probes run every 5 minutes from San Francisco, Ashburn, Montreal, Frankfurt and Singapore. Rank numbers appear once a provider has results for 90% of the week; until then the order is uptime so far.

What we test

ProviderModelAPI hostLinks
OpenAIGPT-5.4 Nanoapi.openai.com
AnthropicClaude Haiku 5.5api.anthropic.com
Google GeminiGemini 3.5 Flash-Litegenerativelanguage.googleapis.com
xAIGrok 4.20api.x.ai
MistralMinistral 3Bapi.mistral.ai
DeepSeekDeepSeek Flashapi.deepseek.com
CohereCommand R7Bapi.cohere.com
Moonshot AI (Kimi)Kimi K3api.moonshot.ai

How we measure

Request
A 1-token chat completion with our own API key
How often
Every 5 min from 5 regions: San Francisco, Ashburn, Montreal, Frankfurt, Singapore
Counted as down
At least 2 regions fail on 2 runs in a row. Why confirm
Not counted
Rate limits (429), errors with our own key or billing, and outages of our own probes
Speed
Request to reply, minus DNS, connect and TLS. Compared only on the same model
Week
Monday to Sunday UTC, frozen when it ends. Ranked with 90% of the week measured

FAQ

Why does this differ from the vendor status page?

We test one API endpoint with one small request; vendors report on many products and regions with their own rules, so both can be right.

Which model do you call?

For LLM APIs, each provider's lowest-cost model. Inference providers all serve gpt-oss-120b or GLM-5.3, so their speed is comparable.

Can a short outage be missed?

Yes. Confirmation takes two 5-minute runs, so an outage shorter than about 10 minutes may not count.

Monitor your own model API with the same check: your key, your model, alerts when it fails.

Start free