Deepgram vs Whisper: features, limits and which to use

How do Deepgram and the Whisper API compare?

Whisper is an open speech model from OpenAI, sold as the whisper-1 API and also hosted by others. Deepgram is a speech company with its own Nova models. The table compares Deepgram Nova-3 with OpenAI’s whisper-1 on facts from each vendor’s pages, checked November 10, 2026.

What does hosted Whisper cost for 100 or 1,000 hours?

Whisper is cheapest when someone else runs it on fast hardware. Groq hosts Whisper Large V3 Turbo at $0.04 an hour. The table covers pre-recorded audio on OpenAI and Groq, and none of these prices includes speaker labels. Prices checked November 10, 2026.

When does running open-source Whisper yourself make sense?

Whisper’s code and weights are MIT licensed, so the model is free to run. You pay for the GPU and the work of operating it. The turbo model is an optimized large-v3 that needs about 6 GB of video memory and runs about 8 times faster than large, with a small loss in accuracy.

Self-hosting fits steady, high volume, strict data rules or offline use. For bursty or small volume, a hosted API at $0.04 to $0.36 an hour usually costs less than keeping a GPU running. Deepgram also offers self-hosted containers, on Enterprise plans with NVIDIA GPUs.

What replaces the Whisper API after February 2027?

OpenAI has deprecated whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, with shutdown on February 26, 2027. For new transcription it points to gpt-transcribe, at $0.0045 a minute. That model takes context such as keywords and expected languages, and can stream partial text while it processes a finished file.

whisper-1 is the only OpenAI model that supports timestamp_granularities for word and segment timestamps. Teams that use Whisper for subtitles should test a replacement, such as Deepgram or a hosted Whisper, before the shutdown.

Can an agent use Deepgram or Whisper through Locus Pro?

Locus Pro offers Deepgram as deepgram/transcribe, paid per second of audio from one prepaid credit balance with no Deepgram account. Nova-3 is the default model, and model=whisper-large runs Deepgram’s hosted Whisper Large. Diarization costs nothing extra; summaries and other options add to the price of the call. Locus has no OpenAI Whisper endpoint. Catalog checked November 11, 2026.

The endpoint takes a public audio URL and fits occasional transcription from an agent. For steady, high volume, your own Deepgram account and API key can cost less per minute.

• Transcribe audio cheaply
• Deepgram pricing in detail

Should you choose Deepgram or Whisper?

Neither vendor page offers an independent accuracy benchmark, so run both on a sample of your own audio before you commit. Accents, noise, jargon and language mix change results more than any headline number.

• Live captions, voice agents and call audio: Deepgram, which streams and labels speakers.
• Meetings with several speakers: Deepgram Nova-3, where diarization costs nothing extra.
• Lowest price per hour, no speaker labels needed: Whisper Large V3 Turbo on Groq.
• Strict data rules or offline work: open-source Whisper on your own GPU.
• Existing whisper-1 code: migrate before February 26, 2027, to gpt-transcribe or another host.
• Occasional transcription from an agent with no Deepgram account: Deepgram through Locus Pro.