Intelligent transcription with Gemini 3.5 Transcribe

Gemini 3.​5 Transcribe entered public preview on August 26, 2026, and it is the first Go​ogle speech-to-text model that is genuinely sold as a G​emini model: you call it through the G​emini API with a model ID, the audio is priced in tokens, and Go​ogle quotes the cost per minute of audio on its pricing page. That last detail is what the consumer coverage glides past. The launch-day stories are about filler-word removal and voice editing in Gboard, but what actually changed for developers is that transcription now sits on the same API surface as the rest of the G​emini line, with speaker attribution and word-level timestamps in the base model rather than in a separate add-on. This piece reads like an API reference, because that is the gap the first wave of coverage leaves open.

Two caveats up front so the numbers below do not mislead. Go​ogle does not call Gemini 3.​5 Transcribe "the most accurate speech-to-text model we have" — the claim is the most precise model for intelligent voice interactions, which is a different claim, and the difference matters when you read the WER figures. And everything here describes a public preview: the two model IDs, the prices, and the benchmark numbers are what Go​ogle ships today, and all of it can move before GA.

The two API paths — and the model IDs to remember

Gemini 3.​5 Transcribe ships as two endpoints, and picking the right one is the first architectural decision a team makes:

• gemini-3-5-transcribe — the pre-recorded path via the Interactions API. Send a file, get the transcript back with speaker attribution (up to three speakers; three-plus is experimental) and word-level timestamps. This is the batch and async workhorse.

• gemini-3-5-transcribe-live — the real-time path via the Live API. Bidirectional streaming with sub-second latency, aimed at live conversation, captioning, and voice-driven interfaces.

The split mirrors how the rest of the G​emini Audio family is arranged, and it matters because "live" is a distinct SKU with its own price. Several early reads of this launch assume one model does both; it does not.

What "intelligent transcription" actually means

Go​ogle's launch post frames this as moving beyond phonetic accuracy toward understanding. The features that follow from that framing are the ones to evaluate rather than the framing itself:

• Disfluency handling — "ums" and "ahs" are dropped, and self-corrections are resolved. "Let's meet Tuesday — no, Wednesday" transcribes as the corrected day, not as a transcript of a false start. That is a decision the model makes, which is the sense in which Go​ogle calls it intelligent.

• Auto-formatting — output comes back as polished text: punctuation, casing, and paragraph structure applied, not a raw token stream.

• Multi-speaker identification — up to three speakers attributed in pre-recorded audio with word-level timestamps; three-plus speakers is experimental. This sits in the base model rather than a separate diarization service.

• Custom vocabulary — you can bias recognition toward jargon, product names, and unusual spellings by supplying a vocabulary, which is the mechanism for vertical domains.

• 85+ languages — automatic detection, with regional accents and dialects called out explicitly.

• Function calling — the model can delegate work such as image generation or file analysis to other G​emini models, which turns transcription into a component of a larger agent rather than a terminal step.

• Entity capture — Go​ogle claims strong performance on noisy, real-world audio and on alphanumeric entities such as postal codes and order IDs.

Press coverage has also reported a 96,000-token context window (roughly an hour of meeting audio without splitting) and emotion detection. Both are consistent with the launch framing, but the model card Go​ogle published does not headline them, so treat them as reported rather than confirmed until a GA spec sheet appears.

The accuracy numbers, labeled

The WER figures Go​ogle cites come from Artificial Analysis: an average word error rate of 4.0% for streaming and 2.6% for non-streaming transcription. On the FLEURS multilingual benchmark, Go​ogle reports 5.50% WER for streaming and 5.04% for non-streaming. Go​ogle also claims a 70% improvement in time-to-final-transcription over its predecessor, Chirp 3.

The honest reading: these are independent measurements in the sense that Artificial Analysis is not Go​ogle, but they are Google-selected measurements, published inside Go​ogle's own announcement, of a model that has been callable for a day. Nobody outside Go​ogle has yet run an adversarial head-to-head against the open-weight models most teams compare it with, and the launch materials contain no direct comparison against the Whisper family or NVIDIA's Parakeet line. The 70% faster-than-Chirp-3 figure is a vendor claim, period. Treat 2.6% and 4.0% as strong early signals, not settled facts.

What it costs

Go​ogle's pricing page quotes each model in tokens and in estimated minutes of audio. For gemini-3-5-transcribe, the blended estimate is roughly $0.005 per minute of audio at list prices — input about $0.003/min, output about $0.002/min. For gemini-3-5-transcribe-live, the blended estimate is about $0.009 per minute. Both carry a free tier today.

Those per-minute figures are estimates derived from an assumed token rate (25 audio tokens per second in, 175 text tokens per minute out), so they shift if Go​ogle changes the token accounting. What is solid is the principle: the list price is the price. That is exactly the principle a pass-through router such as OrcaRouter operates on — 0% markup, provider list price passed through — so if Go​ogle cuts transcription pricing during the preview-to-GA window, the cut is live on our side the same day, not after a reseller updates a rate card.

Where it is rolling out

For developers, public preview today means two surfaces: the G​emini API in Go​ogle AI Studio, and the G​emini Enterprise Agent Platform for enterprise workloads. On the consumer side, Gemini 3.​5 Transcribe already powers the Rambler dictation feature in Gboard on Android and the G​emini app on macOS. Chrome is next — talk-to-type in any web field — followed by Search Live, G​emini Live, Docs, Keep, and Gmail. For a team evaluating it, the practical answer is simpler: it is callable from code today, in English, in the regions where the G​emini API is available.

What this means for your pipeline

The genuinely new thing is not a WER number. It is that Go​ogle is now selling transcription with the two features teams previously assembled themselves — speaker labels and word-level timestamps — in the base model, on a G​emini API surface that supports function calling. If you were building a diarized meeting-notes pipeline out of a plain transcriber plus a separate diarizer plus a formatting pass, this is the first Go​ogle model that collapses those steps. The open question is how the 2.6% non-streaming WER holds up on your own audio, and until an independent reproduction appears, the responsible move for a production team is to run a real sample through the API rather than budget against Go​ogle's chosen benchmarks. For the LLM work that runs on top of the transcript — summarization, structured extraction, action items — that is the layer where routing earns its keep, and it is the layer OrcaRouter covers: one API across 200+ models, automatic failover across providers, and a routing DSL that composes several models into a single call. The transcription step itself you should test with your own audio before you trust anyone's headline number.