Google launched Gemini 3.5 Transcribe on August 26, replacing Chirp 3 as its dedicated speech-to-text model and making it the first STT model sold natively through the Gemini API. Two variants are live now in public preview: gemini-3.5-transcribe
for pre-recorded audio, and `gemini-3.5-transcribe-live`
for sub-second real-time streaming. Both are free to test today in Google AI Studio. If you’re building anything with audio — voice agents, meeting transcription, multilingual apps — this changes your options.
Not Just Transcription — It Has a Reasoning Layer #
Stop thinking of this as a Whisper competitor. The real story is that Gemini 3.5 Transcribe doesn’t just convert audio to text — it reasons about what was said. It strips filler words (“um,” “uh”), handles self-corrections instead of transcribing them verbatim, auto-formats output, and identifies up to three speakers with word-level timestamps. Google calls this “intelligent transcription,” and for once the marketing isn’t wrong.
The feature that actually matters for developers building agents is function calling. Gemini 3.5 Transcribe can delegate tasks to other Gemini models mid-pipeline — a single voice command can trigger image generation, document analysis, or any downstream Gemini action. That collapses at least one layer from most voice agent stacks. Instead of routing audio through STT, then feeding the text to an LLM, then orchestrating tool calls, the transcription model starts the chain directly. It also supports 85+ languages with automatic detection and custom vocabulary bias for domain-specific jargon.
Two Models, Two Jobs, One Price Structure #
The batch model (gemini-3.5-transcribe
) is built for pre-recorded audio. It processes roughly one hour of audio per API call — a 96,000-token context window — and returns speaker-attributed, word-level timestamped output. This is the model you want for meeting analytics, call center reviews, and interview transcription. According to independent pricing analysis from OrcaRouter, it costs approximately $0.005 per minute of audio (about $300 per 1,000 hours).
The live model (gemini-3.5-transcribe-live
) is designed for real-time voice applications — voice agents, live captioning, interactive voice interfaces — with sub-second latency via bidirectional streaming. Cost is roughly $0.009 per minute. For context, Deepgram Nova-3 and AssemblyAI Universal-3 sit in a similar range ($0.006–0.009/min). Whisper, if you’re running it yourself, costs only your hardware. The free preview in Google AI Studio gives you a no-commitment way to benchmark both models against your actual audio before committing.
Where It Wins, Where It Doesn’t #
Google reports 2.6% word error rate on batch and 4.0% on streaming — improvements on Chirp 3, with 70% faster time-to-final-transcription. However, these are Google-selected benchmarks using Artificial Analysis measurements. No independent adversarial comparison against Whisper Large v3 Turbo (2.1% WER on LibriSpeech clean) or AssemblyAI Universal-2 exists yet. Reading self-reported WER comparisons from any vendor is like reading press releases: treat them as directional, not definitive.
The developer community’s reaction was candid. On Hacker News, one developer noted it “definitely needs a bit more work on latency, which is the most important factor for STT apps.” German, Italian, and English mixed-language meetings still see Voxtral Mini performing better in practice. Polish, and some other low-resource languages, favor Whisper Large v3 in community tests. The diarization cap at three speakers also frustrates teams transcribing panel discussions or large calls. And the hallucination risk that plagued Chirp 3 hasn’t been officially addressed — something to verify before shipping to production. The developer discussion on Hacker News is worth reading before you commit.
The decision framework is straightforward. Choose Gemini 3.5 Transcribe if you need speaker attribution built-in, are already building in the Gemini ecosystem, need intelligent formatting out of the box, or want multilingual support without per-language model management. Stick with Whisper if you need self-hosting, have high-volume English workloads where cost matters, or are working in a low-resource language that Gemini hasn’t proven itself on yet. A detailed Gemini 3.5 Transcribe vs Whisper Large v3 Turbo comparison is available if you want the full breakdown.
What This Means for Your Current Stack #
If you’re using Chirp 3, this is your migration path — and it’s an upgrade, not just a rename. Framework integrations are already live: LangChain, LiveKit, Pipecat, and Vercel AI SDK all support Gemini 3.5 Transcribe now. Chrome browser integration is coming, which will open up web-based voice dictation as a platform use case. The model is already powering Gboard Rambler on Android and the “Speak to Window” feature in the Gemini macOS app, which means it’s production-hardened before you even start testing.
Key Takeaways #
- Gemini 3.5 Transcribe launched August 26 — free now in Google AI Studio, replacing Chirp 3 with two distinct API models for batch and real-time use
- Function calling is the biggest differentiator: voice input can trigger other Gemini model actions directly, collapsing your pipeline
- Batch pricing is ~$0.005/min, streaming ~$0.009/min — competitive with Deepgram and AssemblyAI, more expensive than self-hosted Whisper
- WER benchmarks are Google-reported and not yet independently verified — test on your specific audio before declaring it superior to Whisper
- Latency, 3-speaker diarization cap, and unresolved hallucination questions are real limitations to evaluate for production use