Gemini 3.5 Transcribe ships with a 2.6% word error rate and API pricing around half a cent a minute, undercutting the per-seat transcription subscriptions that sales and support ops teams already pay for call and meeting notes.
Every sales team that records its calls pays for the same thing twice. Once for the tool that turns the recording into a transcript, and again in the time someone spends squinting at that transcript before it goes into a coaching review or a compliance file, because the tool mangled a product name or an order number. An Otter.ai Pro seat runs $16.99 a month, or $8.33 on an annual plan, for 1,200 minutes of transcription. That is the going rate for "good enough" text out of a recorded conversation.
What shipped this week changes the math on the "good enough" part. Google introduced Gemini 3.5 Transcribe on August 26, a speech-to-text model that, as measured by the benchmarking firm Artificial Analysis, posts a 2.6% word error rate on pre-recorded audio and 4.0% in real-time streaming mode. Run through the Gemini API, the pre-recorded version costs roughly half a cent a minute, about $0.005, once you blend the input and output token rates Google publishes. Twelve hundred minutes of audio, the same volume an Otter Pro seat caps out at, runs about six dollars in raw API fees. That is not a rounding difference. It is the price of the underlying transcription layer collapsing toward zero while the SaaS wrapper around it kept its subscription price the same.
What actually shipped
Gemini 3.5 Transcribe comes in two flavors. A streaming version, built for live captioning and voice agents, delivers continuous transcription over a websocket connection with sub-second latency. A pre-recorded version, aimed at meetings, call logs, and what Google's own post calls "post-call analytics pipelines," adds speaker attribution and word-level timestamps on top of the transcript.
The model handles the things that make real conversations messy. It cleans up self-corrections ("let's meet Tuesday, no, Wednesday"), strips filler words, and auto-formats the output instead of handing back a wall of run-on text. It also takes custom vocabulary, so a sales team can bias the model toward its own product names and acronyms instead of hoping the model guesses right. Language coverage is broad: over 85 languages with automatic detection, and speaker diarization for up to three participants, with support beyond that listed as experimental.
Google is also positioning it as a straight upgrade over its previous transcription model, Chirp 3, claiming a 70% improvement in time to final transcription alongside the accuracy gains. Streaming pricing runs a bit higher than the pre-recorded tier, about $0.009 a minute blended, which tracks with the added infrastructure cost of holding a live connection open.
Why this matters if you are not the one writing the code
The buyer who should care about this is not necessarily the one signing the Otter or Gong contract. It is whoever owns the call-QA process, the sales enablement stack, or the support ops function that currently pays a per-seat SaaS fee to get transcripts, then still runs a manual pass to catch the errors that per-seat fee doesn't fully solve for. Gemini 3.5 Transcribe is not a meeting notetaker with a nicer interface. It is the raw accuracy layer, priced like a utility bill instead of a software seat, that a team with even modest engineering resources can now build directly into its own call pipeline rather than renting from a vendor whose main job was reselling that same layer with a UI on top.
That distinction matters for the build-versus-buy conversation every ops leader eventually has. At $0.005 a minute, the economics stop being a rounding error and start being a line item worth building around, especially for teams processing call volume in the tens of thousands of minutes a month, where a per-seat SaaS model was never going to scale linearly with usage anyway.
The honest part
This is a public preview, not a shipped, stable product. Google's own pricing page marks the per-minute figures as estimates derived from a token-based billing model, not a locked-in rate card, and enterprise availability through the Gemini Enterprise Agent Platform is listed as "coming soon" in places. Speaker diarization tops out at three confirmed speakers, with anything beyond that explicitly labeled experimental, which matters if your call-QA use case involves group sales calls or panel interviews. And raw API access is not a replacement for what a tool like Otter actually sells: calendar integration, searchable meeting history, shared team workspaces, action-item extraction, CRM sync. Buying the accuracy layer cheap does not make the rest of that product disappear. It just means someone has to build it, or keep paying for it separately.
The transcription vendors that built their pricing around "good enough" accuracy just watched the floor move. The question for the next year is whether they compete on the interface they've already built, or find out how much of their subscription price was actually just a markup on someone else's model.