Cost is only the first filter
Storage, chunk retries, summaries and human review all sit around the speech request. A low transcription rate creates room for those parts, but it does not excuse an unverified output contract.
BATCH TRANSCRIPTION
MAI-Transcribe-2 vs GPT-Transcribe
A product-focused comparison for long podcast and subtitle work: timing, diarization, throughput and the cost of a publishable pass.
Pricing and availability checked September 5, 2026
The short answer
Choose the contract your workflow can verify.
CastTranscript uses MAI-Transcribe-2 for uploaded media because the current price and transcription-specific output fit a batch publishing workflow. The implementation still probes every provider response: model names do not guarantee that a gateway exposes every timing or speaker field.
Decision pointMAI-Transcribe-2GPT-Transcribe
Product role
Speech recognition for files, subtitles and transcript production
File and Realtime speech-to-text option in the OpenAI model family
CastTranscript role
Default upload and explicit live-refine pass
Comparison benchmark, not a hidden fallback
Published price
$0.10 per audio hour as a limited offer through the end of 2026
Check the current OpenAI pricing page for the selected transcription model
Timing and speakers
Microsoft lists word timestamps, diarization and keyword biasing
Response shape and diarization depend on the selected model and endpoint
Language workflow
Microsoft lists 60 languages plus code switching
Language hints and behavior depend on the chosen transcription model
When fields are missing
Keep verified sentence cues and disclose the capability downgrade
Use the exact API response; do not infer alignment
Do not stitch models together
Running one model for text and another for speakers or timing can create subtitles that look precise while the words and cues disagree. One speech model owns each pass.
Probe the response you receive
A gateway can expose fewer fields than the underlying model. Every job records whether segment timing, word timing and speaker labels were actually returned.
CastTranscript decision
MAI handles batch work. Missing detail stays visible.
Uploads use one MAI pass. If word timing or diarization is absent, the project keeps the trustworthy transcript and sentence cues, disables unsupported exports when necessary, and tells the editor what was not returned.
Try an uploadQuestions, answered
MAI and GPT transcription questions
Which model does CastTranscript use for uploaded files?
The default batch pass uses MAI-Transcribe-2. CastTranscript records the provider and model on each job so the model chain is visible.
Does MAI-Transcribe-2 always return word timestamps and speakers?
Microsoft lists word timestamps and diarization as model capabilities, but gateways can expose different response fields. CastTranscript checks the actual response and reports which data were returned.
Will CastTranscript run Whisper or Muse to fill missing fields?
No. The batch path does not silently combine multiple speech models. It delivers verified fields from the MAI pass and shows a capability warning when detail is missing.
Is the $0.10 per hour price permanent?
No permanent price is assumed. Microsoft described $0.10 per audio hour as a limited offer through the end of 2026, so this page includes its verification date.
Private by default
Test the workflow on a recording you know.
Compare the transcript against the source, correct it in one document and export only the formats supported by its timing data.