We are seeing intermittent latency spikes on a Vapi assistant using OpenAI GPT-4.1 with BYOK.
Example call ID:
01a0851d-6dea-7aab-877d-134166a7b932
Timestamp: 2026-09-09
The call logs show seven conversational first-attempt failures with:
providerfault-model-no-response
attemptNumber: 1
totalAttempts: 2
byok: true
Examples:
  • Turn 42: first attempt failed at 07:43:28.832 UTC; first token arrived at 07:43:32.616 UTC — 5.327 seconds.
  • Turn 53: first attempt failed at 07:44:35.661 UTC; first token arrived at 07:44:41.764 UTC — 7.299 seconds.
  • Turn 56 also failed once and retried, with 7.192 seconds to first token.
The retry logs switch between gpt-4.1 and gpt-4.1-2025-04-14. The request uses temperature 0.3, max_tokens 400, and prompt_cache_retention 24h.
Please investigate:
  1. What upstream condition produces providerfault-model-no-response?
  2. Why does the retry switch between the model alias and dated snapshot?
  3. Are timeout or streaming issues contributing to the delay?
  4. Can you provide the underlying provider response or request IDs?
Vapi’s own logs show prompt caching is working, so this report concerns the failed requests and retries rather than caching.