A recorded interview priced by audio duration is different from a live speech session or an audio-token model. Calculate transcription economics from the source meter, not from an assumed word count.
Pricing snapshot: Vendor first-party documentation checked on 25 October 2026. Original price effective date unknown unless separately declared. Amounts are published list figures, subject to tier, region, currency and eligibility.
A concrete per-minute listing
OpenAI’s gpt-transcribe model page lists a cost of $0.0045 per audio minute. Its documentation describes use with completed audio files, streamed file transcripts and committed turns in Realtime sessions. This unit must remain a duration unit in a financial ledger; it is not the same as a per-million-token text rate or a flat “per recording” fee. Store the literal model ID and endpoint alongside the billed audio duration.
The company’s central price sheet separately lists gpt-4o-transcribe and gpt-4o-mini-transcribe with token fields and estimated per-minute costs. Those estimates are not automatically the contractual meter for every path. Distinguish estimated equivalents from a fixed duration rate.
Calculate an archive workload without false precision
Suppose an archive contains 480 billable minutes and each minute is charged at the posted $0.0045 rate. The illustrative model cost is 480 × $0.0045 = $2.16, before taxes, storage, download fees, retries and later translation. That simple formula does not imply the source charges in whole-minute increments; verify any provider-specific rounding and minimum charge rather than inserting one.
If the team cleans audio, separates speakers and processes overlapping clips, the amount of audio submitted may exceed the duration of unique finished transcripts. Budget from submitted billable duration, not the running time of the final edited episode.
An operational pipeline
A robust production sequence includes upload/stream ingestion, language hints if supported, transcription, speaker review, timestamp validation, personal-data handling and export. Domain vocabulary, accents and background noise can make review necessary even when a model advertises high accuracy. This article reports no independent word-error-rate tests.
If the transcript feeds subtitles, validate timing and spelling in the final media context. ASR output alone does not guarantee broadcast-ready SRT; a separate human QA expense may be material in multilingual work.
Live transcription is not file transcription
The general pricing page lists separate live transcription products with their own per-minute rates, and real-time speech-to-speech uses audio/text token meters. A customer-service voice agent and an overnight podcast transcription job therefore require different cost models. When an integration uses multiple APIs, account for each billed product rather than treating transcription as an included feature of an unrelated subscription.
Track timeouts, partial restarts and duplicate job submissions, because these may increase submitted minutes. Idempotency keys and ingestion hashes can reduce preventable reprocessing.
Privacy, retention and region
Recorded speech may contain identities, private conversations and location details. An inexpensive transcription model does not resolve consent, regional processing requirements or deletion obligations. Review the relevant service’s data retention and training policies separately. If a team needs a specific processing region, do not assume the US-dollar global list price buys it in that jurisdiction.
An article checked on 25 October 2026 cannot certify how the vendor priced the same SKU in earlier years. Keep effective_at null unless a dated notice states otherwise, and attach every line item to an official model-specific source.
Pricing evidence boundary
This cost discussion concerns OpenAI (gpt-transcribe; gpt-live-transcribe; gpt-4o-transcribe). Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- developers.openai.com — pricing (official documentation; accessed 2026-10-25)
- developers.openai.com — gpt transcribe (official documentation; accessed 2026-10-25)