Whisper pricing: API rates, self-hosting and migration

What does Whisper cost?

There are three distinct technical purchasing paths. A consumer dictation app with “Whisper” in its name is a separate product. If it charges a subscription, that fee is separate from model licensing or API billing.

OpenAI's Whisper model documentation retains the $0.006 rate and marks the model deprecated. Its transcription deprecation notice sets the shutdown date. That deadline applies to the hosted API; it does not remove the separately released open-source code and weights.

For an existing whisper-1 integration, 100,000 audio minutes at the listed rate produces $600 in transcription usage. A new integration needs a budget for the replacement workflow rather than a long-term forecast built around a retiring endpoint.

OpenAI transcription pricing after Whisper

OpenAI recommends different models for completed recordings and live audio. Its transcription price table lists the following per-minute estimates:

The name gpt-realtime-whisper does not make it the same service as whisper-1 or the locally run Whisper repository. OpenAI's transcription workflow guide distinguishes these paths and does not recommend gpt-realtime-whisper as the starting model for a new integration.

Streaming text from a completed recording is also different from accepting microphone or call audio as it arrives. A file-transcription price cannot stand in for a live-session budget merely because both return text incrementally.

Migration has a feature cost as well as a usage cost

The same transcription retirement also covers gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, and /v1/audio/translations. Moving to one of those deprecated models does not avoid the deadline.

Changing a model ID alone does not preserve every output capability. The file-transcription compatibility documentation describes specialized outputs tied to deprecated models. A migration budget needs to cover application changes around:

• Timestamps and subtitles: existing consumers of word timing, srt, or vtt output need a compatible output path.
• Speaker labels: code that depends on diarization, the assignment of speech to speakers, needs that capability accounted for separately.
• Translation: the replacement for /v1/audio/translations uses a Realtime translation session and a different endpoint, rather than the same completed-file upload flow.

A lower minute rate does not pay for those changes automatically. Keep one-time migration engineering separate from recurring transcription usage.

Is open-source Whisper free?

Whisper's code and model weights are released under the MIT License, with no per-minute model-license charge. The official Whisper repository provides installation and local inference instructions. Running it yourself moves the service responsibilities into your infrastructure budget.

For a workstation, that budget can include hardware depreciation, electricity, and the time spent preparing audio and maintaining the environment. For a cloud deployment, it includes provisioned compute hours, attached storage, networking, and the application around the inference worker.

The model choice matters. Larger models need more accelerator memory, and throughput depends on hardware, audio, decoding settings, and implementation. The repository's relative-speed figures use English speech on an A100 graphics processing unit (GPU). They do not establish how many minutes your chosen deployment will process each month.

Audio hours and compute hours are different units

A hosted audio-minute rate follows the quantity of audio submitted or processed under the service's billing rules. A rented instance follows how long that instance is provisioned. An idle worker can therefore cost money while producing no transcripts.

For a batch workload, a useful capacity calculation is:

required busy compute hours =
  audio minutes / (60 × measured audio-hours processed per compute-hour)

monthly self-host cost =
  provisioned compute hours × compute-hour rate
  + storage and networking
  + operating labor
  + other deployment costs

Provisioned hours include any paid idle time and spare capacity. Parallel workers add compute hours even when they reduce elapsed processing time. A deadline for a large upload batch can require more capacity than the same monthly volume spread evenly across the month.

This is the missing denominator in many “Whisper is cheaper” calculations: a GPU-hour price has no per-audio-minute meaning until a workload has a measured processing rate and a provisioning policy.

A 100,000-minute monthly cost example

Assume 100,000 minutes of completed recordings per month. At the listed gpt-transcribe estimate of $0.0045 per minute, hosted transcription usage is $450. This is the current recommended file-transcription model, not a Whisper price or a claim that its outputs equal local Whisper outputs.

For a contrasting self-host budget, assume one compute instance remains provisioned for 730 hours at an assumed $1 per hour, plus four hours of operating work at an assumed $100 per hour.

The $1 compute rate is a planning input, not a cloud quote. No hardware or processing capacity is attached to it. The example does not establish that one instance can process 100,000 minutes within the required deadline.

The self-host subtotal excludes storage, networking, monitoring, backups, redundancy, security work, setup, taxes, support, and engineering beyond the four assumed operating hours. The hosted usage figure also excludes your application's storage, integration, and operating work. These are two bounded cost lines, not complete deployment quotes.

Volume changes the API line, but capacity changes the self-host line

For the same completed-file workload type, the hosted usage arithmetic is:

Holding the illustrative $1,130 self-host subtotal fixed gives an equal-cost point of approximately 251,111 audio minutes per month: $1,130 ÷ $0.0045.

That intersection is conditional arithmetic. It is not a production break-even result. If the workload requires a second worker, redundancy, or more operating time, the self-host cost changes. If workers shut down between batches, provisioned hours can fall. Neither effect is captured by holding one monthly subtotal constant.

A defensible self-host saving requires both sufficient capacity and a lower fully loaded cost at the required transcript quality. Monthly volume alone proves neither.

How third-party hosted Whisper is billed

A third-party host sells its own service around a Whisper model. It can use a different model version, inference implementation, request schema, queue, or billing meter from OpenAI's API.

Replicate, for example, hosts an openai/whisper model listing using Whisper large-v3. Its public-model billing rules charge active processing time, with setup and idle time free. Public models use a shared hardware pool by default, so requests may queue.

For time-based hosting, the cost calculation is:

transcription inference cost =
  sum(billable processing seconds) × price per processing second

A displayed estimate for one run cannot be converted directly into a universal audio-minute rate. A short recording and a long recording can consume different processing time; model settings and the host's implementation also affect the run.

If a host bills per request or offers a subscription, the arithmetic changes again. File-length distribution, included usage, minimums, concurrency, and overages determine the effective cost. If a plan carries an “unlimited” label, that label alone does not establish the capacity or service terms for a production workload.

Use the same denominator across the deployment choices: total relevant monthly expense divided by accepted audio minutes. Reprocessing and rejected transcripts consume resources without adding accepted output. That keeps an inexpensive run from looking economical when it creates more correction work.

When each deployment model fits

A hosted transcription API fits when you need a service endpoint and want the provider to operate inference capacity. Usage-based billing makes a small or irregular completed-file workload straightforward to budget. Existing whisper-1 applications also carry a migration obligation before the scheduled shutdown.

Self-hosted Whisper fits when local execution, control over the model environment, or a particular deployment boundary matters enough to own the operating work. It can be economical with suitable hardware and utilization, but the license alone cannot establish the saving.

A third-party Whisper host fits when you want access to a specific Whisper implementation without managing its workers. The host's processing meter, queue behavior, output contract, and service terms remain part of that choice.

For live speech recognition, our speech-to-text pricing guide covers streaming billing and provider differences. A batch Whisper hosting estimate is not a substitute for that live workload budget.

Where Dasha fits in a voice product budget

A transcription endpoint produces text. A voice agent also needs conversation execution, model responses, speech output, and production operations. Depending on the deployment, it can also involve telephony and external integrations.

At Dasha, we provide a managed production platform for conversational AI products. We belong in the runtime decision when your team is building and operating agents. We are not a standalone speech-to-text vendor or a line-for-line replacement for a Whisper transcription endpoint.

Our voice-agent cost guide separates runtime, speech, model, carrier, and operating costs. If the product needs a managed conversational runtime, explore our voice backend.