Gemini 3.5 Transcribe gives developers separate live and file speech-to-text APIs at about $0.009 and $0.005 per minute

Google released two dedicated Gemini 3.5 Transcribe API models on August 26, 2026: `gemini-3.5-transcribe-live` for low-latency bidirectional streaming over the Live API, and `gemini-3.5-transcribe` for recorded audio through the Interactions API. Both automatically detect more than 85 languages and support custom vocabulary biasing. The file endpoint can add speaker diarization and word-level timestamps, while Smart transcription can remove filler words and false starts, resolve spoken self-corrections and format raw speech into readable structured text. Google’s current pricing page estimates a blended paid cost of roughly $0.009 per minute for live transcription and $0.005 per minute for non-streaming transcription.

Speech-to-text is becoming an application primitive rather than a separate specialist service. The new Gemini endpoints let builders use the same provider account for live captions, voice interfaces, meeting/call processing and downstream Gemini workflows while choosing explicitly between latency and richer transcript metadata. The price point is low enough that transcription can be embedded into high-volume features, but the two modes are not feature-equivalent: live sessions trade accuracy and metadata for responsiveness. Builders need to pick the endpoint based on interaction design rather than assuming the live model is simply a faster version of the file model.

`gemini-3.5-transcribe-live` provides continuous bidirectional transcription over WebSockets with sub-second responsiveness, aimed at voice agents, captioning and interactive dictation. `gemini-3.5-transcribe` processes uploaded or recorded audio and can spend more time producing richer transcript metadata. Applications should model them as different serving paths rather than one API with a streaming flag.

Google’s Smart mode removes disfluencies such as filler words and stutters, resolves inline self-corrections and applies punctuation, paragraphing, lists, dates and numeric formatting. That is useful for dictation and notes but materially changes the semantics of the output. When exact evidentiary or conversational wording matters, builders should use verbatim mode instead.

The non-streaming endpoint supports word-level timestamps and speaker diarization, while the live endpoint does not. Google documents up to one hour of file audio in ordinary mode, reduced to 30 minutes when features such as diarization or timestamps are enabled. Live sessions are capped at 10 minutes. These constraints can determine architecture before model quality does.

Both modes accept custom vocabulary biasing for domain-specific words, names and acronyms. Google documents up to 1,000 terms, while noting that smaller targeted lists often work best. This can reduce the need for a separate post-processing correction layer in medical, technical, legal or product-specific transcription workflows, although teams should validate terminology on their own data.

Google’s paid pricing estimates roughly $0.009 per minute for the live endpoint and $0.005 per minute for file transcription using its stated audio/text token assumptions. The file endpoint is therefore both richer and cheaper on the published baseline, while live transcription buys immediacy. Free-tier traffic can also be used to improve Google products, whereas the paid tier is marked as not used for product improvement.

Google cites Artificial Analysis measurements of about 4.0% average word error rate for streaming and 2.6% for non-streaming transcription. Google also publishes FLEURS multilingual results. Those numbers are useful comparison points but should not be treated as a guarantee for noisy calls, specialized vocabulary or code-switching; endpoint constraints and your own audio distribution matter as much as the headline WER.