Been hammering the API for a few weeks now. The pattern is too consistent to ignore.
From Friday evening through Sunday night, generations get noticeably worse. More artifacts, weird instrumentation choices, vocals that sound strained or off. Weekday afternoons? Crisp, usable output. It's not subtle. Feels like they're either scaling back compute resources to save costs, or their queues get flooded and quality of service degrades. Either way, it's a problem for anyone building something serious on this.
Anyone else tracking this or am I just unlucky? 🫠
Just my two cents.
Oh wow, I'm actually glad you posted this. I've been using it for a support chatbot prototype and noticed it gets a bit...derpy on Saturday afternoons. I just assumed it was me or my prompts, but the timing lines up.
The "strained vocals" thing especially. It was giving customer replies that felt slightly off-tone, not quite as polished. I had to re-run a bunch.
Have you considered logging the quality scores alongside timestamps? Might help prove the pattern.
Logging timestamps is smart. I've seen similar patterns with other cloud AI services. The "derpy" description is apt.
It could be degraded service from shared weekend load, or they might be pushing less stable model versions during off-peak maintenance windows. Either way, it creates a real problem for anything production-adjacent.
I'd add logging for prompt/response latency and error codes too. Data is the only way to push back on a provider.
—cp