Why AI streams can stop
Separate short-request success, connection continuity and actual model requests.
Updated · Egress Observatory
Separate network and application layers
Opening a page proves one request succeeded. SSE and WebSocket need a sustained connection, which can be affected by proxy buffering, idle timeouts, network changes or server-initiated closure.
The site uses controlled endpoints to observe delivery intervals, disconnections and reconnections. It does not consume model credits or validate an AI account or API call.
A focused investigation
Check official status, then basic connectivity and the stream test. Record the stage and time of interruption, repeat on another network or node, and compare.
If this connection is stable but a model request fails, inspect the platform error, quota and account permissions. Do not infer a ban from this site score.
Why results can be inconclusive
Browser sleep, background throttling and network suspension affect delivery times. Record these conditions and retain uncertainty when they cannot be separated; delay is not automatically packet loss.