ChatGPTの新全二重音声、思考処理をGPT-5.5にオフロード
openaiが全二重音声モデル「gpt-live」をリリース。聞きと話すを同時に行い、推論はバックグラウンドでgpt-5.5に委ねることで知能ギャップを埋める。
openai announced gpt-live, a new voice mode that uses full-duplex architecture to listen and speak simultaneously, mimicking human conversation patterns with interruptions and backchanneling. unlike previous versions, the model offloads web searches and complex reasoning to the gpt-5.5 model in the background, a move that boosts performance on the browsecomp benchmark from 0.7% to 75.2%. the update is available immediately on ios, android, and web for paying and free users, with api access planned for the future, signaling a shift toward multi-model agent architectures that separate conversational flow from cognitive processing.