Reliability: what we measured
These are our own synthetic measurements, dated, with their limits. They describe what happened in the windows we tested. They are not a promise about future behavior and not a service commitment, and a good result in one window does not carry over to other hours. Last updated 2026-11-02.
Adults only (18+). See the acceptable use rules.
The 220-request test on Super Deal (economical tier)
On 31 October 2026 we sent 220 requests (110 streamed and 110 not streamed), one after another with a pause of at least 2.5 seconds, input from 2016 to 8645 tokens, output limit 400 tokens, about 67 minutes in one window, from one address. The histories were synthetic role-play chats with adult characters and neutral content.
| Measure | Result |
|---|---|
| Requests delivered | 220 of 220 |
| Errors before the first text | 0 |
| Streams cut off | 0 |
| Client timeouts (limit 120 seconds) | 0 |
| Full reply time, median / 95th percentile / longest | 8.5 seconds / 34.0 seconds / 51.8 seconds |
| First text of streamed requests, median / 95th percentile | 3.2 seconds / 9.3 seconds |
| Served after the first upstream route failed and a second answered | 6 of 220 (2.7 percent), all delivered |
"Delivered" means the service returned a reply; it does not mean the reply was complete or good, and the next section lists where it was not. A short unrelated test from another process ran on the same line for about three minutes during this window. One window of about an hour, from one address, says little about other hours.
Where it falls short
- Replies cut at the limit. With the regular model at a response limit of 400, 18 of 220 replies (8.2 percent) ended at the limit. In 15 of those the model's hidden reasoning used the budget and the visible reply stops mid-sentence. The service counts these as delivered; you see a cut-off reply.
- Reasoning instead of a reply. 3 of 220 replies (1.36 percent) contained the model's own reasoning text in place of a character reply; in the 30-request cost sample it was 1 of 30 (3.33 percent). We found them by reading, not by an automatic check.
- Slow finish. In the scene runs below, some replies showed their first text within 4.4 seconds but took 44 to 106 seconds to finish (3 of 10 replies in the regular run, 6 of 10 in the fast run). We have not investigated the cause.
A 10-message scene through SillyTavern
We ran the same 10-message role-play scene against the server of SillyTavern 1.19.0, once per setting below (two runs for the last two). A reply counts as complete only if it has text and ended normally, not at the limit. The test used the server API of SillyTavern, not its web interface, and the scene is short.
| Setting | Complete replies |
|---|---|
| Regular model, response limit 400 | 7 of 10 (3 cut at the limit) |
| Regular model, response limit 1000 | 10 of 10 in two runs |
| Fast alias, response limit 400 | 10 of 10 in two runs |
This is why the connection page suggests the fast alias at 400 or the regular model at 1000. We also checked one reply in Janitor AI in a browser, with the regular model name; see the connection page.
Older and longer observations
- 24-hour test, 23-24 October 2026. On Super Deal (economical tier) 98.82 percent of requests (4782 of 4839) were served over 24 hours, with a synthetic mix of 5 request classes. "Served" means a complete answer was returned, not that the text was good. The requests were mostly short, not long role-play histories, and the gateway has changed since.
- Failure bursts. Our own earlier logs (September 2026) record short windows in which individual upstream routes failed for a while. Those are snapshots, not a constant rate, and we do not state a failure rate for them.
- A longer degradation. In one earlier run (2026-09-14, an older gateway version) of 500 requests the corrected failure rate was 40.6 percent and the degradation lasted the whole 5.5-hour run.
What this means for you
This service has two lines, Standard and Super Deal (economical tier); the tests on this page are of Super Deal, which has no backup route outside its own upstream network (Standard has a third provider outside it, see the pricing page). So on Super Deal during a burst of failures requests can fail until the network recovers; in Janitor AI and SillyTavern you can press retry. Check the prices page for the current price.
Questions or problems: support. Rules: acceptable use, terms of service.