Sample: weekly provider report
Period: 23 Oct 2026 through 29 Oct 2026 (seven complete UTC days). A working sample for five selected providers, regenerated with the site. The paid work is choosing your scope, preparing the feed and supporting delivery; this sample is free.
Generated 30 Oct 2026, 17:45 UTC
Download JSON · Markdown report · Discuss a data pilot
What does this benchmark measure?
This independent AI API latency report contains 44,389 probes from 4 regions for 23 Oct 2026–29 Oct 2026: 40,355 network edge checks and 4,034 model requests. Network TTFB and model TTFT are reported separately, with p50, p95 and sample counts. It does not establish a fastest model or a provider SLA.
Network edge response — TTFB
Probe failures are not confirmed provider outages. Sparse means fewer than 10 observations.
On narrow screens, scroll the table sideways to see percentiles.
| Provider / model | Region | Probes | Failed | Success % | p50 ms | p95 ms |
|---|---|---|---|---|---|---|
| anthropic | ap-tokyo | 2017 | 0 | 100.0 | 227.7 | 306.4 |
| anthropic | eu-hetzner | 2018 | 0 | 100.0 | 197.1 | 302.4 |
| anthropic | sa-east | 2017 | 0 | 100.0 | 192.8 | 239.6 |
| anthropic | us-central | 2019 | 0 | 100.0 | 89.6 | 159.9 |
| ap-tokyo | 2017 | 0 | 100.0 | 66.1 | 128.9 | |
| eu-hetzner | 2018 | 0 | 100.0 | 96.7 | 200.7 | |
| sa-east | 2017 | 0 | 100.0 | 137.8 | 701.4 | |
| us-central | 2019 | 1 | 99.95 | 54.1 | 142.4 | |
| groq | ap-tokyo | 2017 | 0 | 100.0 | 172.0 | 250.2 |
| groq | eu-hetzner | 2018 | 0 | 100.0 | 228.1 | 398.7 |
| groq | sa-east | 2017 | 0 | 100.0 | 232.3 | 283.0 |
| groq | us-central | 2019 | 0 | 100.0 | 120.0 | 188.0 |
| mistral | ap-tokyo | 2017 | 0 | 100.0 | 319.0 | 397.4 |
| mistral | eu-hetzner | 2018 | 0 | 100.0 | 101.0 | 207.7 |
| mistral | sa-east | 2017 | 0 | 100.0 | 253.9 | 316.1 |
| mistral | us-central | 2019 | 0 | 100.0 | 194.3 | 270.7 |
| openai | ap-tokyo | 2017 | 2 | 99.901 | 216.5 | 302.5 |
| openai | eu-hetzner | 2018 | 0 | 100.0 | 230.3 | 400.4 |
| openai | sa-east | 2017 | 0 | 100.0 | 244.9 | 432.7 |
| openai | us-central | 2019 | 0 | 100.0 | 139.9 | 224.6 |
Measured model response — TTFT
Probe failures are not confirmed provider outages. Sparse means fewer than 10 observations.
On narrow screens, scroll the table sideways to see percentiles.
| Provider / model | Region | Probes | Failed | Success % | p50 ms | p95 ms |
|---|---|---|---|---|---|---|
| google / gemini-flash-lite-latest | ap-tokyo | 504 | 65 | 87.103 | 1937.5 | 7161.5 |
| google / gemini-flash-lite-latest | eu-hetzner | 504 | 57 | 88.69 | 1171.1 | 6627.1 |
| google / gemini-flash-lite-latest | sa-east | 505 | 55 | 89.109 | 1302.4 | 6123.1 |
| google / gemini-flash-lite-latest | us-central | 504 | 61 | 87.897 | 1622.0 | 6759.4 |
| groq / openai/gpt-oss-120b | ap-tokyo | 504 | 1 | 99.802 | 1596.6 | 2179.4 |
| groq / openai/gpt-oss-120b | eu-hetzner | 504 | 0 | 100.0 | 802.2 | 1397.0 |
| groq / openai/gpt-oss-120b | sa-east | 505 | 0 | 100.0 | 774.2 | 1291.4 |
| groq / openai/gpt-oss-120b | us-central | 504 | 0 | 100.0 | 1310.8 | 1774.1 |
How to read this report
- Network TTFB measures the public API edge, including expected authentication responses; it does not measure model speed or account availability.
- Inference TTFT is measured only for the provider/model pairs shown. Different models are not a like-for-like provider benchmark.
- Failed probes can reflect test credentials, account quotas or network problems; they are not confirmed provider outages. Series with fewer than 10 probes are marked sparse and do not establish ongoing coverage.
- Success is the fraction of received probes marked ok, not time-based uptime or an SLA. Missing probes are not inferred to be successful or failed.
- Percentiles use successful non-null samples across the whole seven-day window, not averages of daily percentiles. Sorted index = floor((p/100)*(n-1)+0.5).
- Unobserved series are omitted; null latency means no successful timed sample. This sample is a format preview, not a customer deployment.
Source: LLMlatency measurements · Methodology. CC BY 4.0, including commercial reuse with attribution.