Reviews · AUGUST 28, 2026
Gemini 3.5 Transcribe ships at $0.005 per batch minute with 2.6% WER
Google opened its new speech-to-text model to developers on August 26 as two separate endpoints — a batch Interactions API and a streaming Live API — with 85-language auto-detection, three-speaker diarization, and filler-word cleanup baked into the base model.
Google put Gemini 3.5 Transcribe into public preview on August 26, priced at $0.005 per minute for batch audio and $0.009 per minute for streaming. That puts a one-hour recording at roughly $0.30 in batch mode and $0.54 live, which is the actual news buried under the benchmark chart.
The model ships as two endpoints, not one. gemini-3.5-transcribe runs through the Interactions API for pre-recorded audio; gemini-3.5-transcribe-live runs through the Live API for streaming. The batch endpoint handles up to an hour of audio per request (30 minutes if diarization or word timestamps are enabled), supports diarization for up to three speakers, and takes custom vocabulary biasing of up to 1,000 terms, though MarkTechPost notes best results come below 100. The Live endpoint wants 16-bit PCM at 16 kHz mono in 100 ms chunks, caps sessions at ten minutes, and drops both diarization and timestamps to hit sub-second latency.
Artificial Analysis puts the model at 2.6% word error rate on non-streaming audio and 4.0% streaming. On the multilingual FLEURS benchmark, Google reports 5.04% non-streaming and 5.50% streaming across the 85-plus locales the model auto-detects. Google also claims a 70% improvement in time-to-final-transcription over Chirp 3, its previous transcription model.
The interesting design choice is smart mode. Google's documented example rewrites "Um, so for the meeting, I think we should, uh, invite Alice and, wait no, Bob and Carol" as "For the meeting, I think we should invite Bob and Carol." Useful. But per MarkTechPost, smart mode can't be combined with word timestamps or diarization, meaning teams that need both a clean readout and an auditable record have to pay for the transcription twice. As MarkTechPost put it: "A readable summary and an auditable transcript are now two different API calls."
Google's launch post names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents as platforms already carrying the Live API. That partner list, more than the WER numbers, is what signals where this lands: the developer stack around real-time voice agents, not the enterprise contact-center rip-and-replace market Chirp was aimed at.
The caveats are the usual ones for a public preview. Pricing isn't guaranteed at GA. The ten-minute streaming cap forces session-stitching for anything longer. Diarization past three speakers is flagged experimental. And there's no self-hosted option, so anyone with audio they can't send to Google's endpoints is still shopping.
At $0.30 an hour, though, the economics of transcribing every sales call, every discovery meeting, every intake conversation stop being a budget question and start being a workflow question. That's the shift.