Gemini 3.5 and GPT-Transcribe: Pricing and Key Features

Two closely timed launches set the stage for automatic transcription. Google updates its lineup with Gemini 3.5 Transcribe, while OpenAI pushes GPT-Transcribe and its live mode. Between displayed costs, published metrics, and features like diarization, the trade-offs differ.

Known Pricing at OpenAI and Lack of Public Pricing at Google

OpenAI displays detailed pricing for GPT-Transcribe: file transcription is charged at $0.0045 per minute, while audio streaming costs $0.017 per minute. According to OpenAI, GPT-Transcribe is offered at a per-minute rate that is 25% lower than that of the previous version. The comparison table highlights these figures and notes that Google has not published per-minute pricing for either streaming or file processing.

Choosing Based on Speaker Attribution and Timestamping

Gemini 3.5 Transcribe offers built-in speaker attribution, reliable for up to three people, as well as word-level timestamps without requiring a separate model. The table indicates that Gemini allows diarization for up to three speakers, while the standard variant of GPT-Transcribe includes neither diarization nor timestamps; these functions require gpt-4o-transcribe-diarize or whisper-1 for timestamps. In this context, Gemini is presented as the most suitable option for meetings, call logs, and other multi-speaker scenarios, avoiding an additional model call that OpenAI's stack still requires. GPT-Transcribe is highlighted for simple transcriptions with a single speaker or live subtitling without attribution.

Published Performance: WER and Language Coverage

For Gemini 3.5 Transcribe, Artificial Analysis metrics cited by Google indicate a word error rate (WER) of 4.0% for streaming and 2.6% for non-streaming. Google also publishes results on the FLEURS benchmark with 5.50% for streaming and 5.04% for non-streaming. For GPT-Transcribe, OpenAI reports a WER dropping from 40.37% with whisper-1 to 19.27% across Common Voice and 22 languages. The summary table recaps these figures. In terms of languages, Gemini 3.5 announces support for over 85 languages; the table also mentions keyword and language indices across a scope of 22 evaluated languages.

Use Cases: Three-Voice Meeting and Live Subtitling

An example for Gemini 3.5 Transcribe involves transcribing the recording of a three-person meeting: sending audio bytes with a natural language instruction produces a segmented transcription by speaker and timestamped, directly usable by an analysis pipeline. For a live event where latency is a priority, OpenAI's gpt-live-transcribe mode uses a persistent WebSocket connection. Partial transcriptions arrive as delta events while the speaker talks, with each audio fragment added to a buffer and then converted to incremental text as soon as the system deems it confident enough.

Architecture and Variants: Lineages and Recommendations

At OpenAI, Whisper was the initial open transcription model, replaced in March 2025 by gpt-4o-transcribe, the first model built on the GPT-4o architecture. GPT-Transcribe, launched in late July 2026, continues this lineage, with an official recommendation from OpenAI regarding whisper-1, gpt-4o-transcribe, and gpt-4o-mini-transcribe for recorded speech. It is also available in a streaming version, gpt-live-transcribe, for low-latency sessions. Google also segments its offering: gemini-3.5-transcribe-live for streaming, with a target latency under one second via the Live API, and gemini-3.5-transcribe for pre-recorded audio via the Interactions API.

Acceleration, Integrations, and Tracking Features at Google

Google announces a 70% improvement in final transcription time compared to Chirp 3, in addition to a gain in accuracy. Gemini 3.5 is capable of delegating tracking operations such as image generation or file analysis to other Gemini models via function calls, this option being accessible in the macOS application. The addition of custom vocabulary complements this integration-focused offering.