OpenAI Replaces Whisper With gpt-transcribe, Slashing Errors by 52%

OpenAI just shipped two new transcription models in the API: gpt-live-transcribe for low-latency live audio streams and gpt-transcribe for completed file uploads and batch workloads. Both are positioned as the new recommended starting points for any transcription integration, replacing the previous defaults of gpt-realtime-whisper-1 and whisper-1.

The headline improvement is not just raw accuracy, it is contextual awareness. Traditional ASR (Automatic Speech Recognition) models treat every audio chunk in isolation. These new models can take in free-form context about the recording, a list of domain-specific keywords, expected languages, and in the case of gpt-transcribe, prior conversation turns. That shift from stateless transcription to context-informed transcription is what makes these models meaningfully different.

The numbers that matter

OpenAI benchmarked both models across several datasets. Here is how they stack up against their predecessors:

The jump for gpt-transcribe on Common Voice is particularly striking: a drop from 40.37% to 19.27% transcription error rate compared to whisper-1. That is roughly a 52% relative reduction in errors. For gpt-live-transcribe on real-world audio, the improvement over gpt-realtime-whisper-1 is about 18%.

OpenAI also introduced a new internal benchmark called Context Aware ASR, which measures semantic accuracy rather than just raw word matching. This is important because a model that transcribes "two" instead of "2" is technically wrong on word error rate but semantically correct. On that benchmark:

• gpt-live-transcribe went from 38.5% semantic accuracy without context to 44.6% with free-form context provided.
• gpt-transcribe went from 41.6% to 45.2% with context.

Context is the real unlock

Both models accept three types of contextual hints that you can pass at inference time:

• prompt: Free-form text describing the recording, its topic, or setting. Think of it like a system prompt for audio.
• keywords: A list of literal terms that might appear in the audio, like product names, medical terminology, or acronyms. These are hints, not forced outputs.
• languages: A list of expected input languages, useful for multilingual recordings or code-switching (when a speaker switches between languages mid-sentence).

For gpt-transcribe specifically, when it runs inside a Realtime API session or a dedicated transcription session, it automatically uses earlier transcribed turns as rolling context. You do not need to manage that manually.

Here is what a basic call to gpt-transcribe with context looks like:

from openai import OpenAI
client = OpenAI()
audio_file = open("medical_call.wav", "rb")
transcription = client.audio.transcriptions.create(
    model="gpt-transcribe",
    file=audio_file,
    prompt="This is a call between a pharmacist and a patient discussing prescriptions.",
    extra_body={
        "keywords": ["metformin", "lisinopril", "Rx", "refill"],
        "languages": ["en"]
    }
)
print(transcription.text)

Where each model fits

gpt-live-transcribe is built for the Realtime API, connecting over WebRTC or WebSocket to transcribe audio as it arrives. It is the right choice for voice agents, live captioning, call center monitoring, or any application where audio is streaming in real time. gpt-transcribe works through the standard /v1/audio/transcriptions endpoint, accepting file uploads up to 25 MB in formats like mp3, mp4, wav, and webm.

The decision tree is simple:

• Live microphone or call stream? Use gpt-live-transcribe.
• Completed recording or batch job? Use gpt-transcribe.
• Need speaker labels (who said what)? Use gpt-4o-transcribe-diarize instead.
• Need word-level timestamps or SRT/VTT subtitle files? Stick with whisper-1.

Pricing

According to the official pricing page, gpt-transcribe is priced at $0.0045 per minute, which is cheaper than the previous gpt-4o-transcribe at $0.006/minute. gpt-live-transcribe costs $0.017 per minute, the same rate as the old gpt-realtime-whisper it replaces. So you get meaningfully better accuracy at the same or lower price.

What you give up

These models are not a universal drop-in for Whisper. There are real capability gaps to be aware of:

• No word-level or segment-level timestamps (requires whisper-1).
• No SRT or VTT subtitle output (requires whisper-1).
• No speaker diarization, meaning no labeling of who said what (requires gpt-4o-transcribe-diarize).
• No audio translation into English (requires whisper-1 via the translations endpoint).

If your pipeline depends on any of those features, you will need to keep using the older models for those specific tasks. OpenAI is explicit that existing integrations can continue using gpt-4o-transcribe, gpt-4o-mini-transcribe, or gpt-realtime-whisper where supported.

The bigger picture

The ASR space has been heating up fast. Whisper was a watershed moment when it launched as open-source, but its stateless architecture was always a ceiling for production use cases. Competitors like Deepgram, AssemblyAI, and Speechmatics have long offered domain-specific models and keyword boosting. What OpenAI is doing here is bringing that same concept into a GPT-native framework, where context injection works more like prompting a language model than configuring an acoustic model.

The practical implication is that the same skills you use to prompt a text model now transfer to audio. A healthcare app can describe the clinical context. A legal transcription tool can list case-specific terminology. A multilingual customer support platform can declare expected languages upfront. Transcription accuracy becomes something you can tune at runtime, not just at training time. That is a meaningful shift in how ASR fits into application development.