Model collapse mitigation? (lost over $500 on this call)

Support Help
1

Gemini 3.5 flash on a single-prompt agent. The user got confused and started interrupting the agent with “hello?” type inputs. The model then raced and streamed “Hello” 50x in a single message after which the user had to hang up.

Please share stats on LLM collapse for commonly used models and which models you recommend for a balance of speed (TTFT with lowest variance), instruction following, and cost.

While i understand, model collapse and similar failures are inevitable, this unfortunate turn of events cost our business over $500 in lost revenue and potential further complications. I’m seeking advice backed by stats to minimize the likelihood of such events in the future.

call: call_9a5a7cf8b3172f1db69e73f06e0

2

Hi @a11 I’ve informed our team about this and will get back to you as soon as I have an update.

Thank You

3

Hello @a11 What you described is consistent with an LLM degeneration / interruption-loop failure mode. This seems to be a gemini-3.5 model issue. The safest default recommendation from what we’ve seen is to move this workflow to gpt-4.1.

Thank You