GPT-Audio

GPT-Audio, released alongside GPT-Realtime on 28 August 2025, is the model most people using OpenAI's voice stack never hear about. Where the realtime model holds an open connection and interrupts itself like a human speaker, this one takes a complete audio file, thinks, and returns a complete answer through ordinary chat completions — the same call developers already use for text. That makes it the model for jobs where nobody is waiting on the line: analysing a recorded call, summarising a voice message, generating narration. The trade-off runs the other way round from what one might expect, and in this model's favour: it holds 128,000 tokens of context to the realtime model's 32,000, and returns up to 16,384 tokens against 4,096. It takes text and audio in, speaks and writes out, and its knowledge stops on 1 October 2023. On 20 July 2026 OpenAI put it on the retirement list together with seven other voice models; it stops answering on 20 January 2027, with gpt-audio-1.5 named as the replacement.

▸Videos

No videos yet.

▸Reviews