gpt-realtime-whisper

Upstream, the distinction is streaming versus batch. On TheRouter that distinction disappears: both models are served over the same POST /v1/audio/transcriptions request, both return one complete transcript, and neither streams. What actually differs here is price and accuracy — and the trade-off runs opposite to what the names suggest. GPT-Realtime-Whisper's list price is $0.017/min against gpt-4o-transcribe's $0.006/min, so it is roughly 2.8x MORE expensive, while gpt-4o-transcribe typically achieves higher absolute accuracy on pre-recorded audio. Unless you specifically need this model, gpt-4o-transcribe is the better default on TheRouter on both counts. (Figures are OpenAI list prices; the charged rates are on each model's pricing card.)

Not natively, and not on TheRouter either. Natively, OpenAI's GPT-Realtime-Whisper uses the dedicated /v1/realtime/transcription_sessions endpoint (WebRTC or WebSocket) — you manage a persistent connection, stream audio via input_audio_buffer.append events, and handle transcript deltas as they arrive. On TheRouter, this model is served only on POST /v1/audio/transcriptions — the standard REST transcription contract, request in and full transcript out. There is no TheRouter chat/completions route for this model; calling it that way returns a 503.

The delay setting controls the latency/accuracy tradeoff for streaming transcription. Lower values (minimal, low) emit transcript deltas faster but may have higher word error rate because the model has less audio context. Higher values (high, xhigh) wait for more context before emitting text, improving accuracy at the cost of display delay. OpenAI recommends starting with 'low' for live captions and 'medium' for balanced use cases. Test with your actual audio — microphones, background noise, accents, and domain vocabulary all affect the best setting.

Prompt steering is NOT supported for GPT-Realtime-Whisper in GA Realtime sessions. Open the session with a language hint instead (the 'language' parameter). Where prompt steering may become available in future snapshots, OpenAI recommends using short keyword lists (e.g., 'Keywords: metoprolol, atorvastatin, A1C') rather than full natural-language instructions. For production, treat keyword steering as an aid, not a guarantee — continue to manually evaluate names, numbers, dates, and domain terms.

Traditional Whisper (whisper-1) processes audio in chunks and must wait for the full utterance before transcribing — it is not natively streaming. GPT-Realtime-Whisper emits incremental transcript deltas continuously as audio arrives, with tunable delay from 'minimal' (near-instant partial text) to 'xhigh' (more context, better accuracy). For live captioning and streaming applications, the latency improvement is dramatic: users see text appearing mid-sentence rather than waiting seconds for the full turn. For pre-recorded files where latency doesn't matter, gpt-4o-transcribe or gpt-4o-mini-transcribe offer higher accuracy at lower cost.