GPT-Realtime-Whisper

GPT-Realtime-Whisper, published on 7 May 2026, revives OpenAI's best-known audio brand for a different job. The original Whisper was a batch model: you handed it a finished recording and waited. This one transcribes a live stream and lets the developer choose where to sit between latency and accuracy — a caption that appears instantly and is corrected a moment later, or a slower one that arrives already right. It is billed by the clock at 1.7 cents per minute of audio, half the price of the translation model in the same family, and runs only on OpenAI's realtime transcription endpoint. The context is small by design, 16,000 tokens with a 2,000-token output ceiling, because live transcription never needs to hold much history. Note that the open-weight Whisper large-v3 has not been replaced or withdrawn: it remains the free, self-hostable option for recordings that are already complete.

▸Videos

No videos yet.

▸Reviews