Google Releases Gemini 3.5 Transcribe with 2.6% WER and Dual Endpoints

Google has released Gemini 3.5 Transcribe, a new speech-to-text model designed for real-time voice interfaces and recorded audio. [1] The system provides automatic detection for more than 85 languages, which includes support for mid-sentence code-switching. [1] Performance metrics measured by Artificial Analysis show an average word error rate of 2.6% for non-streaming tasks and 4.0% for streaming tasks. [1] Furthermore, the model improves time to final transcription by 70% over the previous Chirp 3 model. [1]

The model ships as two distinct endpoints: `gemini-3.5-transcribe` for pre-recorded files and `gemini-3.5-transcribe-live` for bidirectional streaming. [1] These two endpoints do not share the same pricing, limits, or feature sets. [1] The deployment is strictly API-only as a managed service, with no open weights available for self-hosting. [1] Currently, the developer and enterprise tracks remain in public preview. [1]

Google is targeting the enterprise transcription market by splitting Gemini 3.5 Transcribe into discrete batch and streaming endpoints. The 70% speed improvement over Chirp 3 positions it strongly for real-time customer experience applications. However, the lack of an open-weights release forces organizations into a managed-service ecosystem, contrasting with self-hosted open alternatives. What the sources don't address: How the specific pricing structures and rate limits compare between the live and batch endpoints.