Google splits its new transcription model in two — and the pricier half does less

Google made Gemini 3.5 Transcribe generally available on 26 August 2026: its first speech-to-text model sold in its own right rather than as a side capability of a general Gemini model. It ships as two endpoints, and the difference between them is larger than the shared name suggests. The file endpoint, gemini-3.5-transcribe, accepts recordings of up to an hour, detects the language utterance by utterance across more than 85 of them, follows a speaker who switches language mid-sentence, separates up to eight voices and puts timestamps on individual words. The streaming endpoint, gemini-3.5-transcribe-live, works over WebSockets and returns interim results before final ones — but cannot name speakers, cannot time individual words, and holds a session for ten minutes. The prices run the other way round. On Google's own blended estimates the restricted live endpoint costs about $0.009 per minute of audio, the fuller file endpoint about $0.005. In list terms that is $3.50 per million audio input tokens against $2.00, and $21.00 per million text output tokens against $12.00. The buyer of the leaner product pays roughly eighty per cent more. There is a defensible reason — holding an open socket and returning partial text within milliseconds is a different engineering problem from grinding through a stored file — but Google does not offer one, and the documentation states the asymmetry without comment. Anyone planning a live captioning product should know before writing a line of code that speaker labels are simply not on the menu at that price. Two further admissions are worth lifting out of the tables, because vendors rarely print them. Turning on speaker separation or word-level timestamps cuts the maximum recording from an hour to thirty minutes. And word-level timestamps, Google writes, degrade transcription accuracy. Speaker attribution beyond three people is labelled experimental. Custom vocabulary accepts up to a thousand terms, but the maker says results are usually best with no more than a hundred. Both versions have a free tier, and the pricing page states plainly what that tier costs: free-tier requests are used to improve Google products, paid-tier requests are not. Neither version supports function calling, code execution, caching or the thinking mode of Gemini's general models. Polish is among the supported languages.