Since July I've been running probes in four places (Germany, US Central, Tokyo and São Paulo) that connect to the official API host of every AI inference provider I track, every five minutes. That's 44 providers today and 4.8 million measurements in the current 90-day window. Here is what the data says as of November 7, 2026.
Which AI API is the fastest?
It depends on where you call it from. Fastest endpoint per region, median time to first byte over the last 24 hours:
| Region | Fastest API | p50 TTFB |
|---|---|---|
| Germany | Fireworks | 16 ms |
| Tokyo | Fireworks | 19 ms |
| US Central | Fireworks | 30 ms |
| São Paulo | OpenRouter | 62 ms |
The same endpoint can be ten times slower from another continent:
| Provider | Fastest from | Slowest from | Spread |
|---|---|---|---|
| Fireworks | 16 ms, Germany | 254 ms, São Paulo | 16.1× |
| Nebius | 33 ms, Germany | 466 ms, Tokyo | 14.2× |
| SambaNova | 30 ms, Tokyo | 371 ms, São Paulo | 12.3× |
| Upstage | 58 ms, Tokyo | 612 ms, São Paulo | 10.5× |
| GLM | 94 ms, Tokyo | 542 ms, São Paulo | 5.7× |
TTFB here is the network floor: a request over an already open TLS connection until the first byte comes back. It says nothing about model speed or quality.
How much of time to first token is just the network?
Between 2% and 78%, depending on the model and the region. That comes from real streamed completions for the providers where I hold a key. Groq serving gpt-oss-120b has a median TTFT of 354 ms from Germany and 1,371 ms from Tokyo. Same model, same provider, the difference is distance. For a small fast model, picking a host on the right continent matters more than picking the model.
Which AI providers are the most reliable?
Over the last 24 hours, 36 of 44 providers answered every probe (50,006 probes). Widen the window and the picture changes: 26 of 44 dropped at least one probe in the last 7 days. Availability quoted without a window length doesn't compare to anything.
Observed availability over the 90-day window, rounded to 0.1% (what an outside client saw, not an SLA):
| Provider | Observed, 90 days |
|---|---|
| Anthropic | 100% |
| 100% | |
| Groq | 100% |
| Fireworks | 100% |
| OpenRouter | 100% |
| Cerebras | 99.9% |
| SambaNova | 99.4% |
| Doubao | 99.3% |
| OpenAI | 99.2% |
An outage counts against a provider only if it failed from two or more regions in the same minute while no other provider was failing. A failure seen from one region only is listed but never counted.
Does availability depend on where you call from?
Yes. Identical requests to identical endpoints over the last 7 days:
| Region | Failure rate |
|---|---|
| Germany | 0.089% |
| US Central | 0.119% |
| São Paulo | 0.126% |
| Tokyo | 0.174% |
Tokyo failed 2× more often than Germany. A monitor that probes from one location can't see that. Also worth knowing before you plan a fallback around a status page: only 11 of the 44 providers publish a status feed a program can read.
How much notice do providers give before retiring a model?
Computed from 233 records that carry both an announcement and a shutdown date on the provider's own deprecation page:
| Provider | Median notice | Shortest notice |
|---|---|---|
| OpenAI | 183 days | 3 days |
| Mistral | 92 days | 0 days |
| Anthropic | 63 days | 60 days |
| Cohere | 0 days | 0 days |
Coming up next: claude-opus-4-5 retires on Nov 24 and claude-sonnet-4-5 on Nov 30 (replacement: claude-sonnet-5-5). On Dec 11 OpenAI retires gpt-5, gpt-5-mini, gpt-5-nano and o3 (replacements: the gpt-5.6 family).
Caveats
- Probes run from data centres, so absolute numbers are lower than what a home connection sees.
- TTFB is reachability and network distance, not throughput or output quality.
- TTFT is measured only where I pay for a key, so coverage is thin.
Full method: https://llmlatency.dev/methodology
Use the data
Everything is free under CC BY 4.0:
- JSON: https://llmlatency.dev/api/rankings.json, https://llmlatency.dev/api/incidents.json, https://llmlatency.dev/api/deprecations.json
- MCP endpoint for agents: https://llmlatency.dev/mcp
pip install llmlatency- Dataset on Hugging Face (https://huggingface.co/datasets/llmlatency/llm-latency-tracker) and Zenodo (https://doi.org/10.5281/zenodo.23388429)
I built and run https://llmlatency.dev, self-funded, no provider pays for anything. Which providers or regions would you add, and where is the method wrong?
