Gemini 3.5 Transcribe Is Live: It's Already in Gboard, Gmail and Chrome, With a 2.6% Error Rate (2026-08)
Article last updated：2026-08-27
Yesterday (2026-08-26) Google’s official blog quietly dropped a technical post announcing a new speech-to-text model: Gemini 3.5 Transcribe.
That could easily read as just another “the model got better again” headline. What makes this one different is that it’s already rolled into apps you probably have on your phone right now — Gboard, Gmail, Google Keep, Docs — not something you need to sign up for or wait on. This article covers the official numbers, exactly which products have it today, and what it costs developers.
1. What Google actually said
The official blog post (blog.google, 2026-08-26) positions Gemini 3.5 Transcribe as a model for “precise and intelligent real-time transcription.” The key difference from traditional speech recognition isn’t just accuracy — it’s that the model automatically strips filler words, cleans up self-corrections mid-sentence, and auto-formats punctuation and paragraphs, essentially smoothing out the rough edges of how people actually talk.
Official numbers published in the same post (ibid.):
| Metric | Number |
|---|---|
| Word Error Rate (WER, measured by Artificial Analysis) — streaming | 4.0% |
| WER — non-streaming (pre-recorded audio) | 2.6% |
| FLEURS public benchmark — streaming | 5.50% |
| FLEURS public benchmark — non-streaming | 5.04% |
| Latency improvement vs. previous Chirp 3 model | 70% faster “time to final transcription” |
| Languages supported | 85+, with automatic detection |
| Speaker identification | Up to 3 speakers with timestamps on pre-recorded audio (3+ is experimental) |
A lower WER (Word Error Rate) means more accurate transcription; 2.6% roughly means 2–3 mistakes per 100 words transcribed — close to professional human transcription quality.
2. Which products already have it?
This is the part that’s most relevant to you. Per the official post and reporting from 9to5Google (2026-08-26), Gemini 3.5 Transcribe is already live (not a future promise) in:
- Gboard’s “Rambler” voice-typing feature on Android
- The Gemini app on macOS
- Google Antigravity’s prompt-box microphone
- Google’s own post also lists Search Live, Gemini Live, Docs, Keep, and Gmail as rolling it out
Coming next: Chrome browser support for voice typing into web form fields — Google’s post says “coming soon” without giving a firm date.
In other words: if you use Gboard’s voice typing on an Android phone, or dictate emails in Gmail or notes in Keep, this new model may already be transcribing for you with zero setup required.
3. What does it cost developers? We checked the official pricing page directly
The announcement post itself doesn’t list pricing, so this site checked Google’s official AI for Developers pricing page directly (ai.google.dev/gemini-api/docs/pricing, verified 2026-08-27), confirming two corresponding API models:
| Model | Use case | Free tier | Paid tier (audio input) |
|---|---|---|---|
gemini-3.5-transcribe-live | Real-time streaming transcription | Free | US$3.50 per million tokens, or US$0.005/minute |
gemini-3.5-transcribe | Pre-recorded audio transcription | Free | US$2.00 per million tokens, or US$0.003/minute |
Both models have a free tier available for testing — developers don’t need to pay upfront. Google marks this as a public preview stage, accessible via the Gemini API in Google AI Studio; enterprise users go through the Gemini Enterprise Agent Platform.
4. How does it compare to transcription tools already fact-checked on this site?
This site has already fact-checked several dedicated speech-to-text and meeting-transcription tools — here’s how Gemini 3.5 Transcribe stacks up against them:
- Otter.ai: free tier gives 300 transcription minutes per month, capped at 30 minutes per session (per this site’s fact-check), built around meeting notes and summaries as a standalone subscription.
- Descript: free tier gives 60 media minutes per month, with only 100 AI credits total and no monthly refill (per this site’s fact-check), built around audio/video editing paired with transcription.
- Gemini itself: the free tier requires nothing but a Google account, and this Gemini 3.5 Transcribe upgrade is built into existing Google products rather than a new standalone app you subscribe to.
This site’s read: tools like Otter.ai and Descript earn their keep with polished meeting notes, speaker labels, and export formats. Gemini 3.5 Transcribe is positioned more as a quality upgrade across Google’s whole product family’s voice input, not a direct replacement — if you just need voice-to-text on your phone or dictating a quick Gmail reply, this upgrade lands for free with zero effort; if you need full meeting records and post-production, dedicated tools like Otter.ai or Descript still do the heavier lift.
5. What this site couldn’t verify — stated plainly
- Neither the official blog post nor 9to5Google’s report gives a concrete global or region-by-region rollout timeline — both only say it’s “rolling out.” This site does not speculate on when it’ll reach any specific device.
- Whether iOS Gboard or the iOS Gemini app get the same upgrade: the official post specifically names Android Gboard and the macOS Gemini app, with no mention of iOS. Not addressed — stated as unverified rather than assumed.
- No language-by-language accuracy breakdown for Traditional Chinese or any other single language was published; Google only gives the aggregate “85+ languages, auto-detected” figure with no per-language WER numbers.
Sources:
- Google’s official blog — Intelligent transcription with Gemini 3.5 Transcribe (2026-08-26)
- 9to5Google — Gemini 3.5 Transcribe report (2026-08-26)
- Google AI for Developers — Gemini API pricing page (verified 2026-08-27)
Further reading: Gemini free tier fact-check · Otter.ai plans guide · Descript free tier guide · NotebookLM free tier
