Walk me through latency budgets: a feature needs sub-500ms p95. Which model and architecture would you choose at OpenAI?
