← Research

Introducing Omi Medical-1 Flash

Medical transcription for long recordings, at $0.18 an hour. Built on Omi Medical-1 weights, with a faster path from audio to text.

60 minrecorded audio
3.89 smedian processing time
Dedicated RTX PRO 6000 GPU benchmark, six runs with one request at a time. p95: 4.02 seconds. API calls add upload and queueing time.
$0.18
per audio-hour
38% less
than flagship batch pricing
1.04%
medical-term error rate

More recordings. Less waiting.

Flash brings medical speech recognition to recorded consultations, dictation and audio archives. It uses the same medical model weights as Medical-1, with a runtime optimized for speed.

That trades some accuracy for speed and cost. Omi Medical v1.1 remains our most accurate model, and the choice for realtime, speaker labels, custom vocabulary or EU processing. Flash starts with English transcription through asynchronous jobs, processed in the US.

Medical accuracy, measured on the same board.

Flash places second in our batch API table. It ties Mistral Voxtral Small Transcribe 3 on medical-term errors; its higher dose / unit match rate breaks the tie. These are results on Omi’s 1,513-recording development benchmark, not an independent clinical validation.

US production run, 5 November 2026. Same recordings and scorer as the other models on our benchmark.
MeasureFlash result
Medical-term error rate1.04% · 30 / 2,872
Word error rate5.98%
Drug-term errors3 / 442
Dose / unit matches91 / 94

Compare every model and read the methodology.

From five minutes to an hour.

Dedicated RTX PRO 6000. Processing time, not end-to-end API latency.
Recording lengthMedianp95
5 minutes0.75 s1.30 s
30 minutes2.35 s2.42 s
60 minutes3.89 s4.02 s

One request at a time on a dedicated GPU with 16 vCPUs. Ten runs for the 5-minute file; six each for 30 and 60 minutes. Upload, network and queueing add time in the hosted API. These results are separate from the cross-provider speed comparison.

Keep uncertain passages visible.

When Flash flags an unclear passage, the transcript includes [unclear mm:ss-mm:ss] and the response sets review_required. Longer recordings are more likely to include a flagged span.

An unflagged transcript can still contain errors. Review transcripts before clinical use.

Try Flash on your recordings.

Use omi-medical-1-flash with the async jobs API. Your 25 free audio-hours each month are shared across models; Flash costs $0.18 per audio-hour after that.

English only. US processing. The global endpoint routes Flash jobs to the US; the EU-only endpoint does not offer Flash. See the docs for the upload and result workflow.