Speech-to-text model powered by GPT-4o.
GPT-4o Transcribe is OpenAI's next-generation speech-to-text model, released on March 20, 2025. Built on the GPT-4o multimodal architecture, it replaces the legacy whisper-1 API with significantly improved word error rate (WER) across 99+ languages. The model leverages GPT-4o's general language understanding to deliver more accurate transcriptions of technical terminology, proper nouns, mixed-language content, and accented speech — areas where Whisper historically struggled with hallucinations.
The model is available through the existing /v1/audio/transcriptions endpoint with per-minute pricing ($0.006/min), making it cost-effective for high-volume use cases. GPT-4o Transcribe sets a competitive benchmark against industry leaders — Artificial Analysis places it as a top-tier contender alongside Speechmatics and AssemblyAI, trailing ElevenLabs Scribe by just one percentage point in combined accuracy. OpenAI released a December 2025 snapshot (gpt-4o-transcribe-2025-12-15) with further WER gains on Common Voice and FLEURS benchmarks. The model is closed-source and only accessible via API; unlike Whisper, no open-weight version is available.
| Type | Rate |
|---|---|
| Audio | $0.0065 / minute of audio |
| Input | $2.70 / 1M tokens |
| Output | $10.80 / 1M tokens |
| Audio input | $6.48 / 1M tokens |
| Estimated | $0.0065 / minute |
request is estimated price per minute
| Release date | March 20, 2025 (initial snapshot); updated December 15, 2025openai.com ↗ | verified |
| Architecture | GPT-4o multimodal foundation with specialized audio-centric pretrainingopenai.com ↗ | verified |
| Endpoint | /v1/audio/transcriptions (OpenAI-compatible SDK)developers.openai.com ↗ | verified |
| Languages | Exact supported-language count is not publicly disclosed by OpenAI; OpenAI documents better language recognition and FLEURS evaluation coverage spanning over 100 languages.openai.com ↗ | unknown |
| Training methodology | Extensive pretraining on audio-centric datasets + reinforcement learning-heavy paradigm for WER reductionopenai.com ↗ | verified |
| Max file size | 25 MB per request (recommend chunking longer audio < 25 min per file)tokenmix.ai ↗ | verified |
| Supported audio formats | mp3, mp4, mpeg, mpga, m4a, wav, webmdevelopers.openai.com ↗ | verified |
| License | Closed-source / proprietary (API-only access; no open-weight release)techcrunch.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
FLEURS (multilingual WER — all languages) OpenAI reports lower WER than Whisper v2 and Whisper v3 across all FLEURS language evaluations. FLEURS spans over 100 languages; OpenAI does not publish a single aggregate GPT-4o Transcribe score in the announcement. | Outperforms Whisper v2/v3 across all language evaluations per OpenAI | openai.com ↗ | |
Artificial Analysis AA-WER (combined rank) Secondary coverage cites an Artificial Analysis ranking for GPT-4o Transcribe, but OpenAI does not publish an official AA-WER rank or combined score for this model. This page therefore does not treat the secondary ranking as a verified vendor benchmark. | — | Not publicly disclosed | — |
Common Voice & FLEURS (Dec 2025 snapshot) OpenAI says the December 2025 transcription snapshot delivers lower WER than prior models on Common Voice and FLEURS without language hints, and roughly 90% fewer hallucinations than Whisper v2 plus roughly 70% fewer than previous GPT-4o-transcribe models in an internal noise evaluation. OpenAI does not publish the full numeric Common Voice/FLEURS score table for this model. | Lower WER than prior models; exact numeric deltas not publicly disclosed | developers.openai.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/audio/transcriptions -H "Authorization: Bearer $THE_ROUTER_API_KEY" -F "model=openai/gpt-4o-transcribe" -F "file=@speech.mp3"Transcribe audio files with GPT-4o Transcribe through the OpenAI-compatible /v1/audio/transcriptions endpoint via TheRouter.
curl https://api.therouter.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-F "file=@recording.mp3" \
-F "model=openai/gpt-4o-transcribe" \
-F "response_format=json"OpenAI rolled out updated snapshots for gpt-4o-transcribe (2025-12-15) with improved accuracy on Common Voice and FLEURS benchmarks. The update delivers lower word error rates, particularly in zero-shot language detection scenarios. The legacy whisper-1 and initial 2025-03-20 snapshots are expected to retire by June 2026.
OpenAI introduced next-generation speech-to-text models powered by GPT-4o architecture, replacing the legacy whisper-1 endpoint. The models deliver significant WER reductions across 99+ languages, with RL-heavy training reducing hallucinations and improving accuracy on accented speech, technical terminology, and noisy environments. Note: not open-sourced unlike Whisper.
GPT-4o Transcribe uses the GPT-4o multimodal architecture instead of Whisper's dedicated encoder-decoder. It achieves lower WER across 99+ languages (especially on accented speech and technical terminology), is less prone to hallucinations, and uses per-minute pricing ($0.006/min vs Whisper's token-based pricing). However, it is closed-source and API-only — unlike Whisper (MIT license, self-hostable). Through TheRouter, you can access it at competitive routing prices.
No, the base gpt-4o-transcribe model does not identify different speakers. It returns a single transcript without speaker labels. For diarization, use the gpt-4o-transcribe-diarize model (available through TheRouter as openai/gpt-4o-transcribe-diarize) or a third-party service like AssemblyAI.
Not natively. The response_format=verbose_json mode returns segment-level timestamps only. For word-level precision, consider Deepgram or AssemblyAI, which offer per-word timestamping as a native feature.
25 MB per request. For longer audio, split into chunks of up to 25 MB or ~25 minutes each. The model does not have a formal batch processing endpoint for transcription, so you'll need to build your own queue for large media libraries.
The model handles 99+ languages, with strongest performance on English, Spanish, French, German, Italian, Portuguese, Chinese (Mandarin), Japanese, Korean, Arabic, and Russian. WER improvements over Whisper are most notable on languages where Whisper previously struggled (e.g., Indic and Dravidian languages like Tamil, Telugu, Malayalam). For languages not covered well, consider specialized STT providers or self-hosted Whisper models.
Yes. Use model=openai/gpt-4o-transcribe with baseURL=https://api.therouter.ai/v1 and the standard OpenAI SDK. TheRouter routes your request to the best available endpoint for OpenAI's transcription API at competitive rates.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com ↗ | 2026-05-28 | verified |
| Architecture | openai.com ↗ | 2026-08-25 | verified |
| Endpoint | developers.openai.com ↗ | 2026-05-28 | verified |
| Languages | openai.com ↗ | 2026-08-25 | unknown |
| Training methodology | openai.com ↗ | 2026-05-28 | verified |
| Max file size | tokenmix.ai ↗ | 2026-05-28 | verified |
| Supported audio formats | developers.openai.com ↗ | 2026-08-25 | verified |
| License | techcrunch.com ↗ | 2026-05-28 | verified |
| FLEURS (multilingual WER — all languages) | openai.com ↗ | 2026-08-25 | verified |
| Artificial Analysis AA-WER (combined rank) | openai.com ↗ | 2026-08-25 | unknown |
| Common Voice & FLEURS (Dec 2025 snapshot) | developers.openai.com ↗ | 2026-08-25 | verified |
| OpenAI ships December 2025 snapshot with further WER improvements | developers.openai.com ↗ | 2026-05-28 | verified |
| OpenAI launches gpt-4o-transcribe and gpt-4o-mini-transcribe | openai.com ↗ | 2026-05-28 | verified |
| How does gpt-4o-transcribe differ from whisper-1? | openai.com ↗ | 2026-05-28 | to verify |
| Does gpt-4o-transcribe support speaker diarization? | tokenmix.ai ↗ | 2026-05-28 | to verify |
| Can I get word-level timestamps? | tokenmix.ai ↗ | 2026-05-28 | to verify |
| What's the maximum audio file size? | developers.openai.com ↗ | 2026-05-28 | to verify |
| Which languages does gpt-4o-transcribe handle best? | techcrunch.com ↗ | 2026-05-28 | to verify |
| Is gpt-4o-transcribe available through TheRouter? | api.therouter.ai ↗ | 2026-05-28 | to verify |