The data — what is free, and what is available on request
Everything published on this site is free to reuse under CC BY 4.0, including commercially, with attribution. There is no signup, no key, and no rate limit to negotiate. We have recorded 2,463,428 measurements since 2026-07-23 across 45 providers and 4 regions.
What is in the dataset?
One row per probe. Six fields, no derived values stored — percentiles and uptime are computed at read time from these rows, so anyone recomputing them from a bulk export gets the same figures we publish.
| Field | Type | Meaning |
|---|---|---|
ts | ISO-8601 UTC | when the probe ran |
provider | slug | which API host was probed, e.g. openai |
region | slug | one of 4 probe locations |
probe_type | network | inference | edge latency (no key needed) or time-to-first-token (needs a key) |
status | ok | error | whether the endpoint answered |
ttfb_ms | integer | milliseconds to the first response byte, or to the first token |
What coverage does it actually have?
Volume without coverage is a vanity number, so here is the breakdown by region and probe type. Inference rows exist only for providers we hold an API key for — that gap is real and is not filled with estimates.
| Region | Probe type | Rows | Providers | From | To |
|---|---|---|---|---|---|
| Asia (Tokyo) | inference | 8,889 | 4 | 2026-07-24 | 2026-09-08 |
| Asia (Tokyo) | network | 608,307 | 45 | 2026-07-23 | 2026-09-08 |
| Europe (Germany) | inference | 8,802 | 4 | 2026-07-24 | 2026-09-08 |
| Europe (Germany) | network | 603,404 | 45 | 2026-07-23 | 2026-09-08 |
| South America (São Paulo) | inference | 8,884 | 4 | 2026-07-24 | 2026-09-08 |
| South America (São Paulo) | network | 608,082 | 45 | 2026-07-23 | 2026-09-08 |
| US (Central) | inference | 8,889 | 4 | 2026-07-24 | 2026-09-08 |
| US (Central) | network | 608,171 | 45 | 2026-07-23 | 2026-09-08 |
What can I use right now, for free?
- Rankings JSON — current p50/p95 and uptime per provider per region, with the timestamp it was generated at.
- OpenAPI description and an MCP endpoint, so an agent can query the data directly.
- A markdown twin of every page (append
.md), plus the whole corpus as text. - An embeddable badge that regenerates itself.
What is not published?
The site shows a rolling 24-hour window. The raw time series behind it — every individual probe, with DNS, TCP, TLS and first-byte timings separated — is retained for 90 days and is not exposed as a bulk download, simply because serving it well is a different piece of engineering than serving a website. If you need it, ask.
What can I ask for?
- Historical bulk export — the raw series for a period, as CSV or Parquet.
- A region we do not measure yet — probes are small; adding one is a day of work, not a quarter.
- A provider we do not track — if it has a public API host, it can be measured.
- Higher-frequency probing for a specific provider or route.
Open an issue at the public repository and describe what you need. Requests that make the public dataset better get done and stay free; work that exists only for one requester is quoted individually.
Does paying change the rankings?
No, and this is the point. Nothing on this site is ranked by commission, and no provider can pay for placement, ordering, or omission — not now and not later. What is on offer is access to data and additional measurement, never a position in a table. If that separation ever stopped being true the numbers would be worthless, including to whoever paid for them.
Method and limitations: methodology · Licence: CC BY 4.0