GPT-Realtime-Translate

GPT-Realtime-Translate, released alongside GPT-Realtime-2 on 7 May 2026, does one job: it listens to speech in one language and speaks it in another without waiting for the sentence to finish. That streaming behaviour is the whole point — a conventional pipeline transcribes, translates and then synthesises, and each stage adds a pause that makes live conversation awkward. This model collapses the chain and is billed by the clock rather than by tokens, at 3.4 cents per minute of audio. Coverage is deliberately lopsided: it understands more than seventy source languages but speaks only thirteen, which suits its intended use — conferences, support lines and broadcasts where many people speak in and a few major languages come out. It runs only on OpenAI's dedicated translation endpoint, holds a 16,000-token context and, like the rest of the family, knows nothing past 30 September 2024.

▸Videos

No videos yet.

▸Reviews