OpenAI's GPT-Realtime-2 guide sharpens voice agents

Hot take: this is more of a practical operating manual than a flashy model announcement, and that makes it more useful for serious builders.

• –Start with `reasoning.effort: low`; the guide frames higher effort as something to earn, not the default.
• –Preambles are treated as a product feature: useful for noticeable work, but harmful when they create fake chatter or delay.
• –The strongest advice is around control surfaces, not style: define when to act, ask, confirm, retry, or escalate.
• –Unclear audio handling is conservative by design: don’t guess, don’t infer missing words, and don’t burn hidden reasoning on noise.
• –Exact entity capture is a core voice problem here; the guide pushes one-at-a-time collection and explicit confirmation for high-precision values.
• –Long-session behavior is about structure, not brute-force context dumping: separate current state, authoritative sources, and background.