OpenAI's GPT-Realtime-2 tops audio leaderboard

This is a meaningful voice-model win because Audio MultiChallenge stresses messy, multi-turn spoken dialogue, not clean demo prompts. OpenAI is signaling that realtime voice is graduating from novelty to a serious agent surface.

• –Audio MultiChallenge measures instruction following, context retention, and handling corrections in real speech, so the result maps better to production voice agents than simple ASR or TTS scores
• –GPT-Realtime-2 adds GPT-5-class reasoning, 128K context, adjustable reasoning effort, and parallel tool calls, which makes it more relevant for task-completing assistants
• –The new stack bundles speech-to-speech, live translation, and streaming transcription, reducing the need to stitch together separate vendors for voice apps
• –Pricing is still premium, so this looks aimed at enterprise voice workflows where reliability and latency justify cost
• –The benchmark framing matters: OpenAI is competing on conversational robustness, not just latency or natural-sounding audio