I’ve noticed that the agent often has longish pauses (10-20 seconds) when producing output recently, even in relatively short chats.
Steps to Reproduce
Make an agent request (often for something which requires a fair bit of output or thinking, such as writing tests or creating new content).
Agent responds, but pauses mid-sentence for some time (e.g. "The user wants ", "Now I see that ") before completing the response.
Expected Behavior
This is minor but annoying - the agent should respond fairly fluidly, rather than starting a sentence and waiting a long time to continue. This seems to be a recent behavior (more common during the last month or so). Feels like some kind of buffering behavior?
Hey, thanks for the report and the request ID, that’s helpful.
What you’re seeing is likely a mix of two things:
Thinking model behavior: Opus 4.5 in your setup uses high thinking mode. While it’s generating, it may pause visible output to do internal reasoning. Those phases produce tokens you don’t see, so you can get mid-sentence gaps like The user wants followed by silence. That’s fairly normal for thinking models, although 10 to 20 second pauses are on the long side.
Server-side latency: we’ve had elevated agent latency recently, which can stack on top of the thinking pauses and make them worse than usual. The team is aware of it. There’s no timeline for a fix yet, but it’s being tracked.
A couple things you can try:
If the pauses bug you, switching to a non-thinking model like Sonnet 4.5 or Sonnet 4.6 should give smoother streaming with fewer mid-sentence gaps.
If you want to stick with Opus, the pauses should improve once latency settles down.
Let me know if it gets a lot worse or if you notice anything else.
Thanks, Dean, will do. In the scheme of things, this is rather less annoying than the infamous ‘taking longer than expected…’ bug mentioned in other threads. Opus is my go-to (much to the detriment of my credit card balance), so I will keep using it and watch for improvements in response time.