OpenAI releases GPT-Realtime-2 for voice agents

This is a meaningful step beyond “better voice” into “voice as an agent runtime.” The real story is not just the model quality, but the combination of reasoning, tool use, and lower-friction session handling.

• –Supports text, audio, and image input, with text and audio output, so it can sit in multimodal voice-agent flows
• –Built around production features like parallel tool calls, async function calls, and recovery behavior, which matter more than raw TTS polish
• –OpenAI is positioning it for support, assistance, education, and multilingual live workflows, not consumer chat novelty
• –Pricing is still premium, so the first adopters will be teams where latency and call success rates justify the spend
• –The benchmark framing suggests the market is moving from “natural-sounding speech” to “reliable spoken reasoning under tool pressure”