OpenAI Whisper-1

openai/whisper-1 · by OpenAI · encoder-decoder-transformer

Pricing — 1 offering(s)

OpenAI (Audio)

About this offering: OpenAI Whisper-1 speech-to-text, $0.006 per audio minute. Deprecated: OpenAI removes it from the API on February 26, 2027 and recommends gpt-live-transcribe or gpt-transcribe instead, so plan new work on a different model.

Audio minutes

$0.006 / minute Current 2023-03-01 → 2027-02-25

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how OpenAI Whisper-1 fits into a cost-aware routing setup

Capability profile

How OpenAI Whisper-1 rates across core capability dimensions, with the task-level evidence behind each rating.

transcription accuracy Moderate

language support Strong

99 languages with automatic language detection. Includes a translation endpoint that transcribes audio directly to English — unique among the major hosted STT APIs.

speaker diarisation Weak

No native diarization. Word-level timestamps are available but speaker segmentation requires external post-processing (e.g. pyannote.audio).

real time streaming Weak

File-upload API only via /v1/audio/transcriptions and /v1/audio/translations. Latency scales with file duration; not suitable for real-time applications.

custom vocabulary Weak

No native custom vocabulary or keyword boosting. Prompt injection (prepending context text) provides limited guidance but is not reliable.

Benchmarks

Benchmark	Score	Config	Source
WER (LibriSpeech clean)	2.7 %	Whisper large-v2 paper result. API-hosted Whisper-1 uses this model checkpoint.	source ↗
WER (LibriSpeech other)	5.2 %	Whisper large-v2 on the noisier LibriSpeech Other split.	source ↗

Operator guidance

Deprecated: OpenAI removes Whisper-1 from the API on February 26, 2027 (use gpt-live-transcribe or gpt-transcribe), so don't build new long-term work on it. Until then, choose Whisper-1 when broad language coverage (99 languages) or direct audio-to-English translation matters. At $0.006/min it is cost-effective for batch workloads, but Nova-3 and Universal-2 both offer better accuracy on difficult audio. Not suitable for real-time streaming; for that use Deepgram.

Use cases

Batch transcription of multilingual audio across 99 languages
Transcription and translation in a single call (audio → English text)
Research and experimentation where OpenAI platform consolidation matters
Clean audio transcription where noise robustness is not a requirement

Limitations

No streaming endpoint — batch only, file upload up to 25MB
Degrades on noisy, accented, or overlapping audio
No native speaker diarization
No custom vocabulary support