Google just shipped Gemini 3.5 Transcribe, and the headline number is genuinely impressive: a 2.6% word error rate on non-streaming audio and 4.0% on live streams, per Artificial Analysis. That’s a 70% improvement in time-to-final-transcription over Chirp 3, the model it replaces. Language coverage spans more than 85 locales with automatic detection and code-switching handled without configuration — the kind of capability that used to require separate models per language family.
But the architecture decision is where teams need to pay attention. Google split this into two endpoints that don’t share feature parity. The Live API delivers sub-second continuous transcription with interim partials, but caps sessions at 10 minutes and doesn’t support speaker diarization or word-level timestamps. The Interactions API handles pre-recorded audio with diarization and timestamps, but you’re limited to 30 minutes of audio once those features are enabled. If you’re building a real-time voice agent, you’re using Live. If you’re building post-call analytics, you’re using Interactions. They are not interchangeable.
The verbatim versus smart mode split is the design decision most people will miss until it bites them. Verbatim keeps every “um,” false start, and self-correction. Smart mode cleans all of that up — Google’s own example turns “Um, so for the meeting, I think we should, uh, invite Alice and, wait no, Bob and Carol” into “For the meeting, I think we should invite Bob and Carol.” But smart mode can’t be combined with word timestamps or diarization. A readable summary and an auditable transcript are now two separate API calls, not one. That’s a real architectural constraint for compliance-heavy industries like legal and healthcare, where you often need both.
Pricing lands around $0.005 per minute for batch and $0.009 for live, with no open weights and no self-hosted path. This is a managed-service play, not an infrastructure one. The ecosystem is already wired in — LiveKit, Pipecat, Agora, and Vercel all have integrations — which suggests Google learned from the Whisper ecosystem playbook: win the developer tools first, the enterprises follow. Both the developer and enterprise tracks are in public preview, so anyone committing production workloads should treat that as beta-grade stability. The question I’m left with is whether the 10-minute live session cap is a hard technical limit or an early-access constraint that loosens once Google sees real usage patterns.
