Historical snapshot generated on 27 August 2026 from recorded requests. This page is not a live health monitor and does not establish current availability.
| Model | Requests | Completed | Failed before output | Failed during output | Cancelled | Other | Availability per 1,000 | p50 (ms) | p95 (ms) |
|---|---|---|---|---|---|---|---|---|---|
| gpt-4.1 | 6 | 3 | 0 | 0 | 0 | 1 rejected · 2 in flight | 1000 / 1000 | 2677 | 3289 |
| gpt-5.6-sol | 38 | 23 | 12 | 0 | 0 | 3 in flight | 657 / 1000 | 4782 | 6706 |
| gpt-5.6-terra | 3 | 3 | 0 | 0 | 0 | 0 | 1000 / 1000 | 2071 | 2376 |
| o3 | 3 | 3 | 0 | 0 | 0 | 0 | 1000 / 1000 | 1889 | 2564 |
Availability counts completed and cancelled requests against requests that failed before or during output, per 1,000 decided requests, rounded down — so one real failure can never round up to a clean 1,000. Rejected and in-flight requests are counted but decide nothing. Latency percentiles cover only requests that carry a recorded duration; when none does, the cell says so instead of inventing a number. Measurement window: 28 July – 27 August 2026 · 50 request records
The source is the same canonical request log the console reads and billing settles against.
Completed and cancelled requests against requests that failed before or during output, per 1,000 decided requests. Rounding is always down: one real failure never shows as a clean 1,000.
Request duration from receipt to completion, as the system recorded it. p50 and p95 are nearest-rank over the actual sample — every value shown actually happened.
A request with no recorded duration enters no sample, and a window with no requests is shown as exactly that: no requests. Absence is never replaced with a zero or an average.