OpenAI releases GPT-Transcribe API at 3.31% AA-WER and $4.50 per 1,000 minutes

• OpenAI added two transcription models to the API: gpt-live-transcribe for live audio and gpt-transcribe for completed files and batch workloads, according to OpenAIDevs.
• Context is now a first-class ASR input: prompts, keywords, expected languages, and earlier turns lifted OpenAI’s semantic accuracy scores for both new models in OpenAIDevs’ benchmark thread.
• GPT Transcribe landed at 3.31% AA-WER, ranked #9 on Artificial Analysis’ non-streaming leaderboard, while cutting the predecessor price by 25%, according to the AA benchmark thread.
• The migration is mostly a model swap at the endpoint level, but the field changes are real: juberti’s model mapping calls the new models replacements, while OpenAI’s docs keep old models available where supported.

OpenAI’s transcription overview splits the product into two workflows, recorded files and live audio. The migration cookbook buries the useful gotchas: same endpoints, new languages array, JSON response assumptions, streamed file output that is not live microphone input. Artificial Analysis pointed readers to its Speech to Text leaderboard, and juberti left diarization as the comic-relief model slug: gpt-4o-transcribe-diarize.

OpenAI shipped two model IDs, each mapped to a different transcription shape:

• gpt-live-transcribe: low-latency live transcription for microphones, calls, captions, and streaming audio, per OpenAIDevs.
• gpt-transcribe: asynchronous transcription for completed recordings, uploaded files, and batch workloads, per OpenAIDevs.
• gpt-live-transcribe is priced at $0.017 per minute on OpenAI’s model page.
• gpt-transcribe is priced at $0.0045 per minute, or $4.50 per 1,000 minutes, on OpenAI’s model page.

OpenAI’s file transcription guide says gpt-transcribe is the starting model for recorded speech in the original language, with /v1/audio/transcriptions, 25 MB file uploads, and mp3, mp4, mpeg, mpga, m4a, wav, and webm inputs.

juberti described the release as a replacement path: gpt-transcribe replaces gpt-4o-transcribe for offline ASR, and gpt-live-transcribe replaces gpt-realtime-whisper for realtime ASR.

OpenAI’s transcription overview uses softer language: existing integrations can continue using gpt-4o-transcribe, gpt-4o-mini-transcribe, or gpt-realtime-whisper where supported, but they are not the recommended starting models for new transcription integrations.

The context surface is the cleanest API change. OpenAI now exposes three hint types for both new models:

• prompt: free-form context about the topic, setting, or recording.
• keywords: literal names, acronyms, account IDs, medications, product names, or other domain terms.
• languages: multiple expected input languages, replacing the old singular language hint for these models.

Live transcription can also use earlier transcribed turns as context, according to OpenAIDevs’ launch thread. On OpenAI’s Context Aware ASR benchmark, gpt-live-transcribe rose from 38.5% to 44.6% semantic accuracy with free-form context, a +6.1 point gain, while gpt-transcribe rose from 41.6% to 45.2%, a +3.6 point gain.

The realtime transcription guide adds a sharp edge: unsupported language codes reject the request, and keywords reject the whole request or session update if they contain <, >, carriage returns, or line feeds.

OpenAI’s launch numbers and Artificial Analysis’ leaderboard measure different slices, but both moved in the same direction.

OpenAI reported

• Live, Common Voice: gpt-realtime-whisper 20.33% WER to gpt-live-transcribe 19.70%, 0.63 points lower, per OpenAIDevs.
• Live, Real-World Audio Recording: 11.65% to 9.60% WER, 2.05 points lower, per OpenAIDevs.
• File, Common Voice: whisper-1 40.37% to gpt-transcribe 19.27% WER, 21.10 points lower, per OpenAIDevs.
• File, Real-World Audio Recording: 15.21% to 8.98% WER, 6.23 points lower, per OpenAIDevs.

Artificial Analysis reported

• gpt-transcribe: 3.31% AA-WER, #9 on its non-streaming Speech to Text leaderboard, according to the AA benchmark thread.
• Component scores: 2.22% on AA-AgentTalk, 2.72% on VoxPopuli-Cleaned-AA, and 6.10% on Earnings22-Cleaned-AA, according to AA’s dataset breakdown.
• Batch speed: about 34x realtime, far behind Smallest AI Pulse Pro at about 251x realtime in AA’s speed comparison.

AA also said it is still benchmarking gpt-live-transcribe for its streaming Speech to Text leaderboard.

The offline model got cheaper than its predecessor, not cheaper than every OpenAI transcription option.

• gpt-transcribe: $4.50 per 1,000 minutes, according to AA’s pricing thread and OpenAI’s pricing page.
• gpt-4o-transcribe: $6.00 per 1,000 minutes in AA’s predecessor comparison.
• gpt-4o-mini-transcribe: $3.00 per 1,000 minutes, still below the new model in AA’s price chart.
• gpt-live-transcribe: $0.017 per minute on OpenAI’s pricing page, matching the old gpt-realtime-whisper line item shown there.

The launch is better accuracy per OpenAI dollar for the main offline path. It is not the absolute low-cost or absolute high-throughput corner of the STT market.

OpenAI’s migration cookbook says the endpoints stay familiar: file transcription remains POST /v1/audio/transcriptions, and realtime transcription keeps Realtime transcription sessions over WebSocket or WebRTC.

The behavior changes live around fields, output, and workflow boundaries:

• language becomes languages for the new models.
• prompt carries recording context; keywords carry literal terms.
• gpt-transcribe returns JSON, so older text, verbose_json, srt, or vtt assumptions do not carry over automatically.
• File streaming means streamed transcript events for an already supplied file, not an ongoing microphone stream.
• gpt-transcribe can run in a Realtime session after an explicit committed audio turn, over WebSocket.
• Detected language output can be an empty languages: [] array when the model has no reliable prediction.
• gpt-live-transcribe does not return detected-language predictions in its initial contract.

The realtime transcription guide makes gpt-live-transcribe a transcription session, not a spoken assistant session. The example session sets type: "transcription", uses 24 kHz PCM audio, appends chunks with input_audio_buffer.append, and commits turns with input_audio_buffer.commit when automatic turn detection is off.

Applications receive two main event types:

• conversation.item.input_audio_transcription.delta: incremental transcript text.
• conversation.item.input_audio_transcription.completed: final transcript for the committed item.

Latency is now an explicit knob. The guide lists minimal, low, medium, high, and xhigh, with lower delay producing earlier partial text and higher delay giving the model more audio context before emission.

Diarization did not get folded into the new names. juberti said there was “nothing to announce now” on a shorter name for gpt-4o-transcribe-diarize.

OpenAI’s file transcription guide still points speaker-labeled transcripts to gpt-4o-transcribe-diarize with response_format: "diarized_json". The same guide says recordings longer than 30 seconds should use chunking_strategy: "auto", optional known-speaker references are 2 to 10 seconds each, and speaker labeling is available through /v1/audio/transcriptions, not Realtime transcription sessions.