Sending two identical parallel requests is the classic approach. But, logically speaking, it should also double the cost. I would send a second request if the

Sending two identical parallel requests is the classic approach. But, logically speaking, it should also double the cost. I would send a second request if the first request fails to return the first token within, say, 1 second. Then there's a chance the first request is stalling, which is an infrequent event. I wonder if higher-availability tiers of LLM providers do a similar thing internally.