When do you think we can expect cost support for reasoning models? (o1, o3-mini, deepseek R1 via fireworks)