Short answer. Between 18 September and 17 October 2026, the four most available AI APIs we probe were Meta's Llama API (99.986%), Anthropic (99.985%), OpenRouter (99.985%) and Google Gemini (99.983%). OpenAI was at 99.930%. The spread among the big providers is a few hundredths of a percent; the difference that matters shows up in the worst region, not in the average.
These are observed-availability figures, not SLAs: what an outside client saw, from four data centres, every five minutes.
What was measured
- 46 AI inference APIs, probed every 5 minutes from 4 regions: Germany, US Central, Tokyo and São Paulo.
- 1,580,667 probes over 30 full days (18 Sep – 17 Oct 2026, UTC).
- A probe is the network path to the provider's API host: DNS → TCP → TLS → first byte. It fails on a timeout, a connection error or an HTTP 5xx. A 4xx counts as "the API answered" — the probe carries no key.
- Attributed outage = the provider failed from 2+ regions at once and no other provider was failing in the same minutes. A failure seen from one region can't be separated from that network path; a failure shared with other providers has a cause above all of them.
- Probes run directly against each provider, never through a gateway or an aggregator.
Results
| # | Provider | 30-day availability | Worst region | Worst-region availability | Attributed outages |
|---|---|---|---|---|---|
| 1 | Meta Llama API | 99.986% | São Paulo | 99.954% | 0 |
| 2 | Anthropic (Claude) | 99.985% | São Paulo | 99.965% | 0 |
| 3 | OpenRouter | 99.985% | São Paulo | 99.977% | 0 |
| 4 | Google (Gemini) | 99.983% | São Paulo | 99.977% | 0 |
| 5 | AI21 | 99.965% | Tokyo | 99.942% | 0 |
| 6 | Replicate | 99.965% | Germany | 99.942% | 0 |
| 7 | Baseten | 99.959% | Tokyo | 99.942% | 0 |
| 8 | Cohere | 99.959% | Tokyo | 99.919% | 0 |
| 9 | Together AI | 99.959% | Germany | 99.942% | 0 |
| 10 | Groq | 99.954% | Tokyo | 99.907% | 0 |
| 11 | Hyperbolic | 99.951% | Tokyo | 99.885% | 0 |
| 12 | Novita | 99.945% | Germany | 99.872% | 0 |
| 13 | Fireworks AI | 99.939% | Germany | 99.907% | 0 |
| 14 | Perplexity | 99.939% | Tokyo | 99.861% | 0 |
| 15 | Inference.net | 99.934% | Tokyo | 99.885% | 1 (11 min) |
| 16 | DeepSeek | 99.931% | São Paulo | 99.850% | 0 |
| 17 | Mistral | 99.931% | Tokyo | 99.827% | 0 |
| 18 | OpenAI | 99.930% | Tokyo | 99.814% | 0 |
| 19 | Baidu (ERNIE) | 99.928% | São Paulo | 99.792% | 0 |
| 20 | xAI (Grok) | 99.922% | Tokyo | 99.815% | 0 |
| 21 | Reka | 99.919% | Tokyo | 99.815% | 0 |
| 22 | Cerebras | 99.907% | Tokyo | 99.804% | 2 (6 min) |
| 23 | Writer | 99.902% | Tokyo | 99.826% | 1 (7 min) |
| 24 | Sarvam | 99.901% | Germany | 99.849% | 1 (3 min) |
| 25 | SiliconFlow | 99.893% | São Paulo | 99.746% | 0 |
| 26 | Nscale | 99.881% | Tokyo | 99.676% | 0 |
| 27 | Z.ai (GLM) | 99.879% | São Paulo | 99.665% | 0 |
| 28 | Friendli | 99.867% | Tokyo | 99.723% | 3 (5 min) |
| 29 | Typesafe ¹ | 99.852% | Tokyo | 99.712% | 1 (3 min) |
| 30 | Moonshot (Kimi) | 99.850% | Tokyo | 99.792% | 1 (36 min) |
| 31 | Alibaba (Qwen) | 99.826% | Tokyo | 99.676% | 0 |
| 32 | Nebius | 99.812% | São Paulo | 99.607% | 0 |
| 33 | MiniMax | 99.806% | Tokyo | 99.561% | 0 |
| 34 | Targon | 99.803% | Tokyo | 99.561% | 0 |
| 35 | DeepInfra | 99.780% | São Paulo | 99.607% | 0 |
| 36 | Upstage | 99.780% | São Paulo | 99.515% | 0 |
| 37 | Tencent (Hunyuan) | 99.749% | São Paulo | 99.457% | 0 |
| 38 | SenseNova | 99.667% | São Paulo | 99.249% | 1 (2 min) |
| 39 | StepFun | 99.647% | São Paulo | 99.203% | 2 (1 min) |
| 40 | Aleph Alpha | 99.645% | São Paulo | 99.238% | 4 (22 min) |
| 41 | Featherless | 99.642% | São Paulo | 99.434% | 2 (4 min) |
| 42 | SambaNova | 99.535% | Tokyo | 99.260% | 9 (71 min) |
| 43 | iFlytek | 99.266% | Tokyo | 98.406% | 2 (11 min) |
| 44 | ByteDance (Doubao) | 99.260% | São Paulo | 98.891% | 11 (91 min) |
| 45 | Baichuan | 99.243% | Tokyo | 98.394% | 0 |
| 46 | 01.AI (Yi) | 97.275% | Tokyo | 90.650% | 10 (118 min) |
¹ Typesafe was added to the probes on 27 September: 21 days of data instead of 30.
What stands out
- Uptime barely separates the large providers. 24 of 46 APIs stayed at or above 99.9% over the month; the median was 99.90%. Between Anthropic and OpenAI the gap is 0.055 percentage points — about 24 minutes a month.
- The worst region is where the differences are. For 24 providers the worst region was Tokyo, for 17 it was São Paulo, for 5 Germany. US Central was never the worst region for anyone. If your users are in Asia or South America, look at the regional column, not the average.
- Attributed outages concentrate in a few providers. 51 outages in 30 days, across 15 providers; ByteDance Doubao (11), 01.AI Yi (10) and SambaNova (9) account for more than half. The four most available APIs had none.
Limitations, stated up front
- The vantage points are cloud data centres, not consumer networks, so absolute numbers are better than an end user on mobile would see. Comparison between providers is the meaningful part.
- Network-level availability is not the same as a successful completion. A host that answers while the model queue is stuck still counts as up here; inference probes (real time-to-first-token requests) are reported separately.
- 30 days is one window. The live page is rebuilt every 30 minutes from the last 24 hours.
Live data and the dataset
- Current uptime ranking (last 24 h): https://llmlatency.dev/best-uptime
- Per-provider status, e.g. "Is the OpenAI API down?": https://llmlatency.dev/provider/openai
- Incident log with the attribution rule above: https://llmlatency.dev/incidents
- Full daily aggregates, CC-BY-4.0, with a DOI: https://doi.org/10.5281/zenodo.21954788
- From Python:
pip install llmlatency
Disclosure: I run llmlatency.dev, the independent tracker these numbers come from. No provider pays for placement, and nothing on the ranking is affiliate-sorted.
Top comments (0)