Gemini 3.5 Transcribe

Speech-to-text. Audio in, text plus word annotations out, with automatic detection of 85+ languages, mid-session code-switching, speaker diarization (up to 8 speakers), word-level timestamps, smart dictation and formatting, and custom vocabulary biasing of up to 1,000 terms. Status ga because Google's models index badges the Gemini 3.5 Transcribe card 'New Stable' — one card covering both endpoints, gemini-3.5-transcribe and gemini-3.5-transcribe-live — and the id carries no -preview suffix; read 2026-09-01. Paid-tier rates per 1M tokens: input $2.00 on audio, output $12.00 on text. Google also publishes duration-equivalent figures for the same rates (roughly $0.003 per minute of audio in and $0.002 per minute of text out, on an assumption of 25 audio tokens per second and 175 text tokens per minute, blending to about $0.005 per minute), which this schema has no field for. Free tier is free of charge. No context window or max output token count is published for this endpoint, so both fields are null rather than inherited from a sibling; the stated limit is duration — up to 1 hour of audio per request, dropping to 30 minutes when diarization or word-level timestamps are enabled. Caching, code execution, file search, function calling, image generation and thinking are all listed Not supported, as are the batch, flex and priority consumption options. Released August 2026 per Google's deprecations table — month only, which is all Google states, so the field carries 2026-08 rather than a guessed day — and that table says 'No shutdown date announced' with the row not gray. Added 2026-09-01: missing from the catalog since launch, one of four Google models invisible to the check-google-prices.mjs completeness column because all of its price cells are multi-rate (LEARNINGS #105).

Source: Google official documentation ↗

Explore more

Gemini 3.5 Transcribe — questions & answers