Overview
Gemini 3.5 Transcribe is Google’s speech-to-text model for converting prerecorded audio into formatted text. It is designed for uses such as meeting transcription, call-log processing, voice interfaces, and post-call analytics. The supplied model ID identifies the prerecorded-audio model; real-time streaming uses the separate gemini-3.5-transcribe-live model. (blog.google)
Capabilities
- Automatically detects and transcribes more than 85 languages, including language switches within audio.
- Can remove filler words, handle spoken self-corrections, and format transcripts; custom vocabulary helps with specialized terms and names.
- Supports speaker attribution and word-level timestamps for prerecorded audio.
- Google reports a 2.6% average word error rate for non-streaming transcription as measured by Artificial Analysis; actual accuracy depends on the audio and use case. (blog.google)
Limitations
- In Google Cloud’s Gemini Enterprise Agent Platform documentation, word-level timestamps are experimental and can reduce transcription accuracy. Speaker diarization supports up to eight speakers, with attribution for three or more speakers described as experimental.
- On that platform, prerecorded audio is limited to 15 minutes when diarization or timestamps are enabled. Custom vocabulary supports up to 1,000 terms; Google notes that up to 100 terms typically gives the best results.
- The platform documentation lists function calling, context caching, and batch inference as unsupported for this model. Google’s announcement describes function calling as an integration available in the Gemini macOS app, not as a general capability of this model API. (docs.cloud.google.com)
Release and availability
Google announced Gemini 3.5 Transcribe on August 26, 2026, initially as a public preview. Google Cloud’s version documentation lists gemini-3.5-transcribe as generally available from October 2026; it does not specify an exact October release day. (blog.google)
Pricing
Gemini Developer API, standard tier, USD, checked November 16, 2026:
- Audio input: USD $2.00 per 1 million tokens.
- Text output: USD $12.00 per 1 million tokens; the pricing table labels output as including any thinking tokens.
- The listed free tier has no input or output charge, subject to applicable availability and limits.
Google estimates approximately USD $0.003 per minute for input and USD $0.002 per minute for output, or USD $0.005 per minute combined, assuming 25 input audio tokens per second and 175 output text tokens per minute. These are estimates, not separate fixed per-minute rates. The separate live model has different pricing. (aistudio.google.com)