Did you find the estimated latency real or not? I mean the only decent latency choices seems Gpt 4o (or mini) Are the other models really that slower when you try them?