What changed
OpenAI released GPT-Transcribe for completed files and final transcripts of committed Realtime turns, plus GPT-Live-Transcribe for continuous low-latency streaming. Both accept free-form context, keyword hints and multiple expected languages, with different outputs and latency contracts.
Why it matters
Teams can choose accuracy-oriented final transcription separately from live partials instead of treating one speech model as a drop-in path. The split makes endpoint, transport, reconciliation, latency and output-format compatibility explicit migration decisions.
What to do
Replay consented recordings through the old and new paths, measure word and named-entity error by language and noise level, then test partial-to-final reconciliation, delay, disconnect recovery and downstream formats. Keep Whisper where native subtitles, timestamps or translation remain required.
Caveat
Availability and interface behavior are OpenAI evidence; no independent multilingual accuracy or latency result is supplied. The new models are not drop-in replacements for every Whisper response format, timestamp or translation workflow, and live audio adds privacy and transport risks.