Asked about Streaming from the model to the screen

What happens when a user cancels a streamed answer?

An answer that works

The server should notice the client disconnect and cancel the upstream request, so generation stops. Otherwise the model keeps producing tokens you pay for and nobody reads.

Reading an answer is not being able to give one. The section that teaches this is Streaming from the model to the screen, 4 hours of Streaming responses.

Learn this properly

Also asked about Streaming responses

Does streaming reduce latency?What does streaming complicate?Would you stream a response that must be valid JSON?Your streaming chat works locally but the whole answer appears at once in production. Why?What metric does streaming improve?How do you handle a tool call that arrives in a stream?