GPT-Realtime-1.5

GPT-Realtime-1.5, released on 23 February 2026 alongside GPT-Audio-1.5, was OpenAI's mid-generation speech-to-speech model, presented by the company as its best model for audio in and audio out and aimed squarely at voice agents and customer support lines. It accepts text, audio and images and answers in text or speech over the Realtime endpoint, with function calling and prompt caching. What separates it from the 2.x models that followed is context: it holds only 32,000 tokens and returns at most 4,096, which is enough for a support call but not for an agent that has to keep a long session and a pile of tool output in view. Audio costs 32 dollars per million input tokens and 64 per million output, text 4 and 16 dollars, images 5 dollars on the way in. Its knowledge ends on 30 September 2024, and it runs only on /v1/realtime — chat completions, batch and assistants are not available.

▸Videos

No videos yet.

▸Reviews