Yeah that actually makes sense. GPT-5.4 mini is likely more optimized or even specially tuned on Retell’s side, so they’re squeezing extra performance out of that specific model. That’s why you’re seeing a bigger gap there compared to others. The STT spikes you’re noticing are probably because they’re using a more aggressive or less stable speech-to-text setup to keep latency low, and since they don’t expose the provider, it’s harder to control or debug. If you want, I can help you replicate the fast response feel on Vapi while keeping STT more stable it’s doable with the right setup @cem802