$ cat wiki/models/gemini-3-5-transcribe.md
Gemini 3.5 Transcribe
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind |
| Released | 2026-08-26 |
| Announced | 2026-08-26 |
| Context window | unknown |
| Pricing | unknown |
| License | proprietary |
| Availability | Google Antigravity; Gboard Rambler; developer APIs. Coming to Search Live, Gemini Live, Docs, Keep, Gmail and Chrome |
Context window is unknown because nothing read states an audio-length or token | |
limit; Pricing because no rate was published, and no API model id was surfaced | |
| either (source). |
Release Date
2026-08-26. A transcription model with automatic detection of more than 85 languages (source).
The interesting claim is not accuracy, it is that the output is not a transcript. The model adapts unstructured speech into formatted text and removes filler words — it is doing editing, not only recognition. That makes its WER a measure of one of its two jobs, and nothing read measures the other: there is no figure for whether the formatting or the filler-word removal is correct, and no statement of what a user gives up when a disfluency carried meaning.
Benchmarks
Google's own figures, on FLEURS, "across a set of top languages and locales":
| Mode | WER |
|---|---|
| Streaming | 5.50% |
| Non-streaming | 5.04% |
| Against Google's own Chirp 3 (2025): improved word error rates and | |
| "significantly better latency" — time to final transcription improves by 70% | |
| (source). |
Two caveats travel with those numbers. The comparison baseline is Google's previous model and no one else's — no third-party ASR system appears. And "across a set of top languages and locales" is doing real work in a claim about 85+ language detection: FLEURS covers 102 languages, the subset scored is not named, and a WER averaged over the best-resourced ones is not the number a low-resource user gets.
Use Cases
Dictation and note capture is the deployment: Gboard Rambler ships on it, and Search Live, Gemini Live, Docs, Keep, Gmail and Chrome are named as coming. It is available to developers via API, and in Google Antigravity today (source).
Compared To
- SL2T — DeepMind's other 2026-08 recognition model, sign-language-to- text, shipped on Pixel through Gboard and Live Transcribe. Same month, same surface, same pattern of publishing the capability and not the accuracy — except that SL2T published no figure at all, where Transcribe publishes a WER against its own predecessor.
- GPT-Realtime-2 (OpenAI) and Grok Voice Think Fast 2.0 — the speech-to-speech tier. Different task: those models answer, this one writes down. No shared benchmark exists between them in this wiki.