One duration meter
The catalog currently has 50 routes with one usable audio/input-duration meter. Multiplying that same rate by 60 or 3,600 stays inside the provider's listed billing dimension.
Speech-to-text cost guide
Transcription API pricing: compare the billing unit first
A duration rate and an audio-token rate are different billing dimensions. This guide keeps them separate, so an hourly transcription estimate never depends on an invented token-to-minute conversion.
Billing basis
Three cases need different math
Audio-token billing
8 routes currently expose an audio-input token price. Their transcript text-output price is shown separately when present; neither rate is converted to minutes.
Directional or incomplete meters
3 routes expose positive input and output duration meters, while 5 current routes do not fit either safe comparison. These stay outside workload-total rankings.
Duration-priced routes
Compare like-for-like input-duration rates
Sorted by the listed input-duration rate. This is not a claim about total transcription cost, quality, latency, or availability.
| Provider | Model | Billing basis | Per minute | Per hour | 10 hours | 100 hours |
|---|---|---|---|---|---|---|
| Groq | whisper-large-v3-turbo | Listed input-duration meter | $0.0007 | $0.04 | $0.40 | $4.00 |
| Groq | whisper-large-v3 | Listed input-duration meter | $0.0018 | $0.11 | $1.11 | $11.10 |
| assemblyai | best | Listed input-duration meter | $0.0020 | $0.12 | $1.20 | $12.00 |
| Mistral | voxtral-mini-2602 | Listed input-duration meter | $0.0030 | $0.18 | $1.80 | $18.00 |
| elevenlabs | scribe_v1 | Listed input-duration meter | $0.0037 | $0.22 | $2.20 | $22.00 |
| elevenlabs | scribe_v1_experimental | Listed input-duration meter | $0.0037 | $0.22 | $2.20 | $22.00 |
| deepgram | nova | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2 | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-atc | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-automotive | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-conversationalai | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-drivethru | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-finance | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-general | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
| deepgram | nova-2-meeting | Listed input-duration meter | $0.0043 | $0.26 | $2.58 | $25.80 |
Showing the lowest 15 of 50 directly comparable duration-meter rows. Open the full transcription model directory for the rest.
Audio-token routes
Keep audio input and transcript output separate
No minute or hourly equivalent is shown because the catalog does not define a universal conversion from audio tokens to time.
| Provider | Model | Audio input / 1M tokens | Transcript text output / 1M tokens |
|---|---|---|---|
| Azure | gpt-4o-mini-transcribe | $1.25 | $5.00 |
| OpenAI | gpt-4o-mini-transcribe | $1.25 | $5.00 |
| OpenAI | gpt-4o-mini-transcribe-2025-03-20 | $1.25 | $5.00 |
| OpenAI | gpt-4o-mini-transcribe-2025-12-15 | $1.25 | $5.00 |
| Azure | gpt-4o-transcribe | $2.50 | $10.00 |
| Azure | gpt-4o-transcribe-diarize | $2.50 | $10.00 |
| OpenAI | gpt-4o-transcribe-diarize | $2.50 | $10.00 |
| OpenAI | gpt-4o-transcribe | $2.50 | $10.00 |
Directional duration meters
Do not silently add input and output rates
These rows expose positive input and output per-second values. The normalized catalog does not by itself prove that both charges should be summed into one transcription total, so the two directions are displayed separately.
| Provider | Model | Input duration / second | Output duration / second |
|---|---|---|---|
| Azure | whisper-1 | $0.0001 | $0.0001 |
| watsonx | whisper-large-v3-turbo | $0.0001 | $0.0001 |
| OpenAI | whisper-1 | $0.0001 | $0.0001 |
Incomplete comparison data
Some routes need exact-model verification
5 current transcription routes do not expose a safe duration or audio-input-token comparison in the normalized fields used above. They are not treated as free or unavailable.
File vs realtime
Billing data does not classify every workflow
Prerecorded-file transcription and realtime streaming have different product requirements, but a model name is not enough evidence to label a route as one or the other. This guide does not infer workflow class from words such as “live”, “realtime”, “file”, or from a missing streaming flag.
Use the exact model page and its provider source when endpoint or serving-mode details matter. A future catalog row with an explicit reviewed serving mode can be labeled more narrowly without changing this comparison rule.
Use the catalog as the source of the comparison
Pricing rows change over time. The directory gives you the complete current transcription subset, while each model page preserves the exact normalized fields and source links used by this guide.