OpenAI GPT-Realtime-Whisper: Streaming STT Guide

OpenAI spent years letting Whisper be the world's most downloaded speech model while selling almost no managed STT story. That changed with the GPT-Realtime family: GPT-Realtime-Whisper (streaming speech-to-text), GPT-Realtime-Translate (live translation), and GPT-Realtime 1.5 (the voice model itself).

If you're choosing a realtime stack in late 2026, here's where OpenAI actually stands — beyond the Whisper nostalgia.

The Lineup

Model	What it is	Key fact
GPT-Realtime-Whisper	Streaming STT, transcribes live as you speak	Low-latency transcription experiences
GPT-Realtime-Translate	Live speech translation	70+ input → 13 output languages, keeps pace with speaker
GPT-Realtime 1.5	Full voice model (listen + reason + speak)	Successor to the 4o voice stack
GPT-4o Transcribe	File-based STT	$0.006/min, per-minute billing
GPT-4o Transcribe Diarize	STT with speaker labels + timestamps	$0.006/min, same meter
Whisper (open)	Self-hostable baseline	Free; Large v3 still the benchmark reference

Also in the wild: gpt-transcribe / gpt-live-transcribe model IDs inside Microsoft Foundry (July 2026), and the classic whisper endpoint for offline English transcription/translation.

How It Compares to the Realtime Field

No independent daily benchmarks (Coval-style) exist yet for GPT-Realtime-Whisper — it's too new. Evaluate on architecture and fit:

	GPT-Realtime-Whisper	Deepgram Flux	Universal-3.5 Pro RT	Scribe v2 Realtime
Transport	Realtime API (WebSocket)	WebSocket	WebSocket / Agent API	Streaming API
Translation built in	Yes (Translate sibling)	No	No (add-on)	Translation model separate
Reasoning-adjacent	Native (GPT stack)	No	Agent context feature	No
Proven latency numbers	Not yet published independently	Promo-priced, widely deployed	6.99% WER published	123ms finals, slow first token
Price signal	Via Realtime API usage	$0.0065/min EN	$0.45/hr / $4.50 agent API	$0.39/hr

The strategic shape is clear: OpenAI sells transcription as an input to reasoning, not a standalone product. If your agent already runs on OpenAI models, keeping audio in the same API removes a vendor, a bill, and a failure hop.

Strengths

• One-stack realtime: STT → reasoning → TTS → translation inside the Realtime API. Nobody else bundles this tightly.
• Translate is genuinely differentiated: 70+→13 live translation with speaker-pace tracking has no direct single-model rival.
• Diarize variant closes the speaker-label gap for file workflows at the same $0.006/min.

Weaknesses

• No independent accuracy numbers yet. Every competitor above ships with daily leaderboard data; OpenAI ships demos. Insist on a bake-off.
• $0.006/min file pricing is 2–3x dedicated async rates ($0.21–0.29/hr). You pay the OpenAI tax for convenience.
• Ecosystem gravity, same as Google: data retention (30-day abuse monitoring), DPA/enterprise gating for zero-retention.

Verdict

• Pick GPT-Realtime-Whisper if your agent is already OpenAI-native and you value the STT→reasoning handoff over shaving milliseconds or cents.
• Benchmark Flux (turn-taking), Universal-3.5 Pro Realtime (entities), and Scribe v2 Realtime (voice-matched pipelines) before committing — realtime is the one category where vendor demos diverge most from production audio.

FAQ

Does this replace Whisper? No — open Whisper remains the free self-hosted baseline and the "Large v3" reference every benchmark still cites. See our Whisper architecture explainer.

Whisper API vs GPT-4o Transcribe vs Realtime-Whisper? Files in English → classic/4o Transcribe. Files needing diarization → Diarize variant. Live microphone → GPT-Realtime-Whisper.

Related comparisons & reviews

• Deepgram Flux vs. VAD: Why Turn Detection is the New Latency Battleground
• ElevenLabs Scribe v2 Review 2026: Realtime & Pricing
• Gemini 3.5 Transcribe Review: Smart STT Tested (2026)
• Microsoft MAI-Transcribe-1.5 Review: In-House STT