Gemini 3.5 Transcribe: pricing and API limits

Gemini 3.5 Transcribe ships as two models, not one

Google announced the model on August 26, 2026. It does one job, turning audio into text, and is separate from the general-purpose Gemini 3.5 line that also writes and reasons.

Recorded audio and live audio use different model IDs

Audio you already have on disk goes to gemini-3.5-transcribe through the Interactions API. Audio arriving from a microphone goes to gemini-3.5-transcribe-live through the Live API.

Latency is not the only difference. Speaker attribution and word-level timestamps exist on the recorded path alone, so a product that starts on the live path cannot add them later without a second pass over the same audio. Choose the path from what your output has to contain, not from which one sounds faster.

Smart transcription stays off until you ask for it

Two modes exist. The default, verbatim, returns what was said including every "um" and every abandoned sentence. Switching to smart drops the filler, resolves self-corrections so that "Tuesday, no, Wednesday" comes back as Wednesday, and formats lists and dates on the way out.

For a one-person voice memo headed into a written report, smart is the mode you want. For a five-person standup, choosing it costs you the speaker labels, which is usually the whole reason the recording exists.

Recorded audio transcribes more accurately than live audio

Google reports accuracy as word error rate, the share of words a transcript gets wrong, measured by Artificial Analysis. It is not a single number: streaming scores worse than batch, because the model has to commit to words before the sentence finishes.

Google measures accuracy separately for each mode

The gap between the two modes is small in absolute terms and consistent in direction. Batch processing wins on every benchmark Google published, including the multilingual FLEURS set, where the model also improves on the previous Chirp 3 generation.

What changed most is not accuracy but waiting. Time to final transcription improves by 70% against Chirp 3, and that is the number a captioning interface exposes to its users, rather than anything an offline pipeline would notice.

The model detects the language on its own

Language selection is automatic across more than 85 languages, and the model follows speakers who switch languages mid-sentence rather than locking to whichever one it heard first. Passing a language code is still worth doing when you already know the answer, since it narrows what the model has to weigh.

Live transcription costs about twice as much per minute

Live audio is the more expensive path on every line of the table, and the difference compounds with volume rather than staying a rounding error.

Speech to text pricing is published per token and per minute

Recorded audio blends to about $0.005 per minute and live to about $0.009. An hour of audio therefore runs about $0.30 against $0.54, and a thousand hours a month about $300 against $540 — a gap wide enough to decide whether captions run continuously or only when somebody asks for them. A free tier covers both models.

Those per-minute figures are Google's own estimate rather than a metered rate, derived from 25 audio tokens per second of input and 175 text tokens per minute of output. Speech that produces denser text than the assumption will land above the estimate.

Smart transcription cannot run alongside diarization

Enabling smart rules out both diarization_mode and timestamp_granularities. Transcribing a five-person meeting therefore means taking verbatim output with the speaker labels intact, then cleaning it up as a separate step in a different model.

Diarization is only dependable for three speakers

The documentation accepts up to 8 speakers, and the same page marks attribution beyond 3 as experimental. Treating the ceiling as the working figure puts production output on an experimental path, while designing for three keeps you inside what Google says it supports.

Turning on extras cuts the maximum audio length in half

A standard request accepts an hour of audio, and that drops to 30 minutes once diarization or word-level timestamps are enabled. A pipeline built around hour-long meeting files breaks the moment somebody switches on speaker labels, so chunking belongs in the design rather than in a later patch. Live sessions cap separately at 10 minutes each.

Where to start depending on what you are building

Building minutes for a call with five or more people? Start with the Interactions API in verbatim mode. Keeping the speaker labels matters more than readable prose, and the cleanup can run afterwards in a separate model.

Turning a single-person voice memo into a report? Same API, but choose smart. Only one person is speaking, so there is nobody to tell apart.

Putting live captions on a support screen? That is the Live API. Speaker separation is not offered there at all, which takes the decision away.

Pricing for the general-purpose Gemini models is covered in Gemini 3.5 pricing and free tier limits explained. To read Google's announcement as plain text, the tool below converts the page to markdown.