Cache measurements
October 9, 2026. Text-only synthetic requests with a connected account. Observed results; not a guaranteed provider TTL.
| Target pause | Conversation | Observed pause (s) | Input tokens | Cached tokens | Cache share | Request ID |
|---|---|---|---|---|---|---|
| 1m | continued | 60.0 | 20060 | 19456 | 96.99% | chatcmpl-861a7497-2453-4c27-b0ad-046bed1cde5e |
| 1m | fresh | 60.0 | 20046 | 19456 | 97.06% | chatcmpl-9532d697-4762-4627-a5fd-683d5272687b |
| 5m | continued | 300.0 | 20061 | 19456 | 96.98% | chatcmpl-9682c309-fcdd-4c13-97fa-9de37767e1c4 |
| 5m | fresh | 300.1 | 20047 | 19456 | 97.05% | chatcmpl-b0b6ad8d-c3af-4a46-838f-b1605bb2902a |
Longer-duration measurements were stopped. A cache hit depends on the model, eligible prefix and provider routing. These samples do not establish a 24-hour provider cache lifetime. Gateway context retention is a separate capability.