Hosted Whisper, without the GPU bill.
Whisper large-v3 and large-v3-turbo on an OpenAI-compatible API from $0.0012 per minute. Batch and streaming, with diarization, speakers and understanding available on the same request.
Whisper is a great model and an awkward dependency.
Whisper is open-weight, multilingual and genuinely strong on real-world audio, which is why so many products start with it. The difficulty is never the model — it is everything around it.
large-v3 wants real GPU memory. That hardware sits idle between jobs, so steady cost meets bursty demand. Whisper is also a batch model by design: making it feel real-time means chunking, windowing, handling partial results and stitching segments back together, and then keeping that working. And Whisper only transcribes. It has no notion of who is speaking, no entity extraction, no sentiment — so the interesting work still needs another vendor.
What hosting it here changes
You get the same open-weight models with per-second billing instead of reserved hardware, streaming already solved, and the rest of the pipeline available on the same request. Add diarization, speaker identification, formatting, entities or sentiment by naming the stages you want; each is metered on its own, so a plain transcription request costs plain transcription.
Economical without giving up speed
Cheaper hosting normally means a queue you wait behind. We price low by running a lean inference stack rather than by throttling: turbo is the low-latency default, and higher tiers raise concurrency, token processing throughput and latency ceilings. Pro allows 50 concurrent batch jobs and 25 concurrent streams with priority scheduling; Enterprise adds custom limits, dedicated hosting and capacity planning for burst volume.
large-v3 or large-v3-turbo.
What you stop maintaining.
Whisper large-v3 needs real GPU memory and stays idle between jobs. Hosted, you pay per processed second instead of per reserved hour.
A backlog of ten thousand recordings does not need a capacity project. Concurrency scales with your plan.
Whisper is a batch model by design. Running it as a low-latency stream is engineering you would otherwise build and maintain.
Diarization, speaker ID, entities and sentiment attach to the same job, which self-hosted Whisper alone will never give you.
What Whisper alone will not give you.
Split the recording into speaker turns and align them to Whisper output.
Attach turns to enrolled voiceprints when you need real identities.
Punctuation, casing, numbers and paragraphing that stay stable across runs.
People, organizations, places, dates and domain values pulled from the transcript.
Conversation-level or segment-level sentiment scored alongside speakers.
Summaries, extraction and classification through hosted language models on the same job.
Questions, answered.
How much does the Whisper API cost?
Whisper large-v3-turbo is $0.0012 per minute in batch ($0.07 per hour) and $0.0018 per minute streaming. Whisper large-v3 is $0.0018 per minute in batch ($0.11 per hour). Billing is per second with no minimums, and no seat or subscription requirement on pay-as-you-go.
What is the difference between Whisper large-v3 and large-v3-turbo?
Turbo is a distilled, faster variant with substantially lower latency and cost, and is the better default for most production workloads including streaming. large-v3 is the fuller model and is the one to reach for when accuracy on difficult audio matters more than speed or price.
Is this the same Whisper that OpenAI released?
Yes — these are the open-weight Whisper models, hosted by us. You get the same model family with our inference stack, pricing and pipeline around it, rather than a reimplementation.
Is the API OpenAI-compatible?
Yes. The audio endpoints follow OpenAI conventions for routes, Bearer authentication, errors and usage, so migrating usually means changing a base URL and an API key. Ultrafield capabilities are added as extensions rather than by changing the compatible surface.
Should I self-host Whisper instead?
Self-hosting makes sense when you have steady, predictable volume and GPU capacity already. It is less attractive when load is bursty, because you pay for idle hardware, and it leaves you to build streaming, diarization, retries and observability yourself.
Can I use Whisper with diarization?
Yes. Diarization, speaker identification, formatting, entities and sentiment run as stages on the same request, so a Whisper transcript can come back already split by speaker and enriched.
What languages does Whisper support?
The Whisper family is multilingual across a broad set of languages, with quality varying by language and audio conditions. See the models documentation for current coverage and guidance.
Can I switch between Whisper and your own model?
Yes, by changing the model name in the request. Ultrafield ASR-1 is our proprietary engine for difficult audio; Whisper models are open-weight alternatives. Switching engines never changes the integration.
Run Whisper without running GPUs.
Same open-weight models, per-second billing, streaming solved, and the rest of the pipeline one field away.
From $0.0012/min · Batch and streaming