GPT-4o Audio, available since 1 October 2024, was OpenAI's first model to take a sound file in and give speech back through the same chat completions endpoint developers already used for text — no new protocol, no open connection. It reads text and audio, answers in text and audio, holds 128,000 tokens of context, returns up to 16,384, and its knowledge stops on 1 October 2023. It never left preview status: the model ID still ends in -preview, and the default snapshot is dated 3 June 2025. Sound is the expensive part — audio costs $40 per million tokens in and $80 out, sixteen and eight times the text rates on the same model. On 20 July 2026 OpenAI put the whole legacy voice family on the retirement list; GPT-4o Audio stops answering on 20 January 2027, with gpt-audio-1.5 named as the replacement.