GPT-Realtime, released on 28 August 2025, ended the preview era of OpenAI's voice API. For eight months developers had been building on gpt-4o-realtime-preview, a model the company itself labelled experimental; this release was the first the company declared ready for production traffic, and it arrived with the Realtime API leaving beta on the same day. The gains over the preview were measured rather than claimed: 82.8 per cent on Big Bench Audio against 65.6, 30.5 per cent on MultiChallenge instruction following against 20.6, and 66.5 per cent on ComplexFuncBench function calling against 49.7. It also cut audio prices by a fifth, added image input, remote MCP servers, reusable prompts and — unusually for a voice API — SIP, so a model could answer an ordinary telephone call. Its limits show its age: a 32,000-token context window, 4,096 tokens of output and knowledge stopping on 1 October 2023. OpenAI announced its retirement on 20 July 2026, with gpt-realtime-2.1 named as the replacement.
#speech to speech#voice agents#realtime#api only#retiring