OpenAI Whisper / GPT-4o Transcribe

OpenAI's speech-to-text lineup spans three generations under one API: the original open-weight Whisper (released Sept 2022, MIT-licensed, 99 languages, the hosted whisper-1 endpoint), and the 2025 GPT-4o-based transcription models, gpt-4o-transcribe and the cheaper gpt-4o-mini-transcribe, plus the late-2025 gpt-4o-transcribe-diarize variant that adds built-in speaker attribution. All are reachable through the single /v1/audio/transcriptions endpoint (and several over the Realtime API), which makes OpenAI the default low-friction choice for any team already on the OpenAI platform: one API key, one SDK, near-zero integration cost. Whisper additionally remains downloadable and self-hostable, which is rare among the major commercial STT vendors and a major reason it dominates the open-source ASR ecosystem.

On accuracy, the GPT-4o transcribe models are a genuine step up from Whisper for multilingual and noisy audio, OpenAI reports lower word error rate than Whisper v2/v3 across the 100+-language FLEURS benchmark, and independent reviewers cluster gpt-4o-transcribe with Deepgram Nova-3, AssemblyAI and ElevenLabs Scribe inside a 2-5% WER band on standard benchmarks. Notably, on clean read-speech (LibriSpeech test-clean) the older Whisper Large-v3 (~2.7% WER) still edges out OpenAI's own reported ~4.1% for gpt-4o-transcribe, so the GPT-4o models win on robustness and languages rather than on every academic metric. The headline weakness is architectural: because the GPT-4o models are LLM-based, they can "follow" instructions embedded in the audio rather than transcribing them verbatim. Simon Willison and OpenAI's own engineers have flagged this; for sensitive verbatim use cases, plain Whisper is arguably safer.

The practical gaps matter more than the WER deltas for many buyers. The hosted API caps uploads at 25 MB (roughly 25-30 minutes), forcing client-side chunking for long files; word-level timestamps are weaker than purpose-built tools (WhisperX, Deepgram); and true speaker diarization only arrived with gpt-4o-transcribe-diarize, which as of mid-2026 is Transcription-API-only with documented teething issues on Azure and in maintaining speaker identity across chunks. Pricing is simple and competitive at the low end ($0.006/min for whisper-1 and gpt-4o-transcribe, $0.003/min for mini) but token-based GPT-4o billing can surprise on long audio, and there is no published STT-specific SLA. Net: an excellent default for general-purpose, multilingual transcription embedded in an OpenAI-centric stack; less ideal where you need guaranteed verbatim output, long-file handling without chunking, or contractual uptime guarantees.