Google's new transcription model didn't reset the speech-to-text price floor — it finally brought Google down to it

What actually shipped

Two model IDs, both in public preview through the Gemini API and Google AI Studio:

• gemini-3.5-transcribe — file processing, via the Interactions API
• gemini-3.5-transcribe-live — streaming, via the Live API

Google’s stated capabilities are specific enough to check. Automatic detection across more than 85 languages, handling regional accents and dialects, with code-mixing supported mid-session. Smart transcription handles self-corrections, strips filler words, and auto-formats output. Custom vocabulary biases recognition toward domain jargon and unusual spellings, capped at 1,000 terms. Speaker diarization and word-level timestamps for pre-recorded audio. And function calling, which lets the transcription model hand off to other Gemini models mid-task.

It is already shipping inside Google products: Gboard Rambler on Android, the Gemini app for macOS, and Google Antigravity’s prompt box, with Chrome named as next. Enterprises get it in public preview through the Gemini Enterprise Agent Platform.

That is a substantial release. The question is what it does to the market, and the answer is not the one in the headlines.

The comparison that matters

Nearly every writeup of this launch quotes Google’s per-minute figure without loading in the add-ons that competitors charge separately — which is precisely the mistake that makes speech-to-text procurement go wrong. The relevant number is what you pay per audio-hour for pre-recorded transcription with speaker labels, because almost nobody transcribing meetings or calls needs it without them.

Google arrives third, not first. It is roughly 30% more expensive than AssemblyAI and 7% more than Deepgram for the same job.

The streaming picture is slightly worse. Gemini’s live endpoint blends to about $0.009/minute, or $0.54/hour, against roughly $0.45/hour for AssemblyAI streaming and $0.46/hour for Deepgram. On real-time transcription, Google is the expensive option among the three — and, as covered below, the one that cannot attribute speakers while streaming.

None of this makes the release unimpressive. It makes the framing wrong. The speech-to-text price floor was not moved this week. It was set by the specialists well before this launch, and Google has now met it rather than broken it.

The 64% that did happen

Look one row further down that table and the actual repricing is obvious.

Google Cloud Speech-to-Text — the Chirp-based product Google has been selling to enterprises — costs $0.48/hour for batch, and diarization is a $0.36/hour add-on, which is eighteen times what Deepgram and AssemblyAI charge for the same feature. Loaded, that is $0.84/hour.

Gemini 3.5 Transcribe does the same job, more accurately, at $0.30/hour with diarization bundled. That is a roughly 64% reduction in Google’s own transcription pricing, and it comes with the 70% latency improvement over Chirp 3 that Google leads with.

So the story is not Google attacking Deepgram. It is Google retiring an uncompetitive product it had been charging enterprise rates for, and folding transcription into the Gemini platform where the pricing is set by what the market already charges. That is a meaningful event — it is just an internal one, and it tells you something about where the pressure came from. Google did not price this against the specialists to win share from them; it priced it at the floor because that floor is now simply what speech-to-text costs, and continuing to charge $0.84/hour against it had become untenable.

Where the accuracy claim holds and where it breaks

The 2.6% figure is better evidence than most launch benchmarks because it comes from Artificial Analysis rather than Google’s internal evaluation. Independent measurement is worth something, and the pattern of vendors marking their own homework makes third-party numbers genuinely more credible.

What that number cannot do is rank Google against its competitors, because word error rate is close to meaningless across different test sets. The published figures currently circulating — Deepgram at 5.26%, AssemblyAI Universal-2 at ~14.5%, ElevenLabs Scribe v2 at ~3.3%, Google at 2.6% — are measured on four different corpora with different audio quality, accent distributions, domain vocabularies, and disfluency-scoring conventions. Deepgram’s number is deliberately drawn from hard real-world medical, finance and call-centre audio. ElevenLabs’ is English-only internal evaluation. Ranking them produces a leaderboard that means nothing.

The clearest demonstration is inside Google’s own release: the same model scores 2.6% on the Artificial Analysis set and 5.04% on FLEURS multilingual. One model, one week, and the accuracy claim nearly doubles depending on what you play it.

Run your own audio. A few hours of representative recordings, scored against a corrected transcript, will tell you more than every published WER in this category combined.

The limits that decide whether you can use it

These are in the documentation rather than the announcement, and they are the details most likely to disqualify the model for a given application:

• One hour per request for file processing — but 30 minutes when diarization or word-level timestamps are enabled. Since those are exactly the features meeting transcription needs, the practical limit is 30 minutes, and longer recordings require chunking and rejoining.
• 10 minutes per session on the live endpoint. Continuous transcription of a long call means building session rotation and reconnection handling.
• No diarization or word-level timestamps in live mode at all. These are file-processing features only. If you need real-time speaker attribution — live captioning that labels who is speaking — this model does not do it today, and a specialist streaming vendor still does.
• 1,000 custom vocabulary terms maximum.
• A discrepancy to verify: the API documentation says diarization handles up to eight speakers; Google’s announcement blog describes accurate attribution for up to three. Test against your own multi-party audio rather than assuming the higher figure.

Add to that the ordinary preview caveat: this is public preview, not GA, and preview pricing on Gemini models has a history of being introductory. Assume the rate can move before it stabilises.

The billing model nobody is mentioning

Every specialist in this category bills a flat rate per audio-minute. Google bills per token, and publishes a per-minute conversion.

For audio input that distinction does not matter — audio tokenises at a constant rate per second, so a minute of silence costs the same as a minute of dense speech. But the text output side is billed on the transcript you actually generate, at $12.00 per million tokens for batch. Four people talking quickly for an hour produce substantially more output than a sparse one-on-one with long pauses, and you pay the difference.

That variable component is roughly 40% of the blended batch estimate and 44% of the live estimate. Which means Deepgram’s $0.26/hour is a quote and Google’s $0.30/hour is an estimate. On a fixed small volume this is noise. On a large deployment, or anywhere you have to forecast spend rather than simply pay it, it is a real structural difference — and it runs in the wrong direction, because the workloads with the densest speech are usually the high-value ones you most want to transcribe.

What this changes for what you buy

If you buy meeting software, nothing. Otter.ai, Fireflies.ai and Granola do not compete on raw transcription accuracy — they compete on calendar integration, bot join behaviour, searchable cross-meeting history, CRM sync and permissions. Speech-to-text is one upstream component, and when it improves your vendor adopts it and you notice better transcripts. Our audio tools guide is unaffected by this release, and so is the choice between those three.

If you call an STT API directly, your position depends entirely on where you are now:

• On Google Cloud Speech-to-Text — migrate. Better accuracy, 70% lower latency, and roughly 64% lower cost. This release is aimed at you.
• On AWS Transcribe — the loaded-cost gap is more than five-fold. Worth a serious evaluation even accounting for egress and integration work.
• On self-hosted or API Whisper — Gemini is cheaper than the Whisper API and adds diarization, which Whisper cannot do. The main reason to stay is data residency or an air-gapped requirement, which remains a legitimate reason.
• On Deepgram or AssemblyAI — stay, unless testing on your own audio shows a real accuracy gain on your domain. Price is not your reason to move, and for live transcription with speaker labels the specialists are still the only ones who can do the job.

If you are building voice products more broadly, this sits alongside the realtime voice model consolidation as the same underlying trend: the audio layer of the stack is being absorbed into the general-purpose frontier platforms, and the standalone-component vendors are being pushed toward the parts the platforms have not bothered with yet. Real-time diarization is currently one of those parts. It will not stay one indefinitely.

The read

The interesting thing about Gemini 3.5 Transcribe is not that it is cheap. It is that Google could not price it above the specialists even with a frontier lab’s accuracy and its own distribution into Chrome, Gboard and Android.

Transcription has become genuinely commoditised. The floor is set by companies far smaller than Google, it did not move this week, and the largest AI company on earth arriving with better accuracy still had to meet it rather than beat it — while cutting its own legacy product by nearly two-thirds to do so.

For buyers that is good news of a specific kind: the price of speech-to-text is now stable, competitive, and low enough that it should not be the deciding factor in what you build. Pick on latency, on session limits, on whether you need speakers labelled in real time, and on what your own audio actually scores. Those still differ meaningfully between vendors.

The price no longer does.

Questions or corrections? Email Pick Right. Want the full list? See all news.