Gemini 3.5 Transcribe Live vs Gemini 3.5 Transcribe: which endpoint your app should call

When Go​ogle launched the Ge​mini 3.5 Transcribe family on August 26, 2026, it shipped two models that share a name and a mission but are built for different stages of a conversation: Ge​mini 3.5 Transcribe Live, the streaming endpoint served through the Live API, and Ge​mini 3.5 Transcribe, the pre-recorded endpoint served through the Interactions API. The mistake most teams make in the first week is treating this as one model with two buttons. It is two SKUs — different APIs, different prices, different hard limits — and the accuracy comparison between them is not a fair fight, because they are measured under different rules. This piece sets the two side by side so the decision turns on your workload, not on a leaderboard.

The two at a glance

• API — Ge​mini 3.5 Transcribe Live runs on the Live API (WebSocket, bidirectional streaming) vs Ge​mini 3.5 Transcribe on the Interactions API (file in, transcript out).

• Job — live conversation, captioning, voice agents vs meetings, call logs, archival audio.

• Accuracy — 4.0% WER streaming vs 2.6% WER non-streaming, per Artificial Analysis, cited by Go​ogle; different measurement regimes, not directly comparable.

• Latency — a final transcript 0.40s after speech ends vs batch processing of a complete file.

• Session — 10-minute cap per live session vs up to 60-minute files (30 minutes with diarization and timestamps enabled).

• Speakers — none on the streaming side vs up to three speakers with word-level timestamps on the pre-recorded side.

• Price — about $0.009 per minute blended vs about $0.005 per minute blended.

The architecture decision: Interactions API or Live API

Both models sit under Go​ogle's Ge​mini Audio family, and both are in public preview. The difference starts at the transport. gemini-3-5-transcribe-live opens a WebSocket and keeps it open: raw audio streams in continuously — the common framing is 16 kHz or 24 kHz 16-bit PCM chunks — and the model returns partial transcripts while you speak, then a final transcript for each utterance when the speech ends. gemini-3-5-transcribe takes a complete file, processes it, and returns the finished transcript with speaker attribution and word-level timestamps.

The transport decision drives everything else. If your product renders words as they are spoken — a live caption, a voice agent reading intent mid-sentence, a call assistant acting while the call is live — you are on the Live path. If your product produces a record — a meeting note, a call log, an archive — you are on the Interactions path. The confusion is that both produce text; the difference is whether the text is a real-time state or a final artifact.

Accuracy: the streaming tax is real

Artificial Analysis, the independent benchmark house whose numbers Go​ogle cites in its launch post, measures a 4.0% average word error rate for the streaming endpoint and 2.6% for the non-streaming one. On the FLEURS multilingual benchmark, Go​ogle reports 5.50% streaming against 5.04% non-streaming. The 2.6% figure puts Ge​mini 3.5 Transcribe at #5 on AA's WER leaderboard at launch, behind ElevenLabs Scribe v2 at 2.2% and Microsoft's MAI-Transcribe-1.5 at 2.4% — those are AA's measurements, not Go​ogle's claims.

The streaming number is not a worse version of the same test. It is a different regime: the model commits to words before it has heard the sentence end, and it pays for that. Do not pick between the two on accuracy alone, because on any given workload the gap could flip. Pick on the constraints: the streaming endpoint has a 10-minute session cap, no speaker diarization, and no timestamps; the pre-recorded endpoint handles hour-long files, attributes up to three speakers, and returns word-level timestamps. If your app needs speaker labels, the streaming model simply cannot do the job, whatever its WER.

What they share

The two endpoints share the "intelligent transcription" engine, and on the shared features the choice between them is neutral:

• 85+ languages with automatic detection, including mid-stream language switching on the Live path.

• Disfluency cleanup — fillers dropped, self-corrections resolved ("Tuesday — no, Wednesday" lands as the corrected day).

• Auto-formatting — punctuation, casing, and paragraph structure applied to the output.

• Custom vocabulary — up to 1,000 terms for jargon and product names, with Go​ogle's docs recommending roughly 100 for best results.

One caveat that applies to both: the "smart" cleanup that makes output readable is a design decision that trades verbatim fidelity. Awkward phrasing gets smoothed; the text is the model's reading of intent, not a courtroom record. For exact transcripts you will want a verbatim option where one is exposed, and you will want to verify the behavior on your own audio either way.

Cost per minute: the live premium

Go​ogle prices audio in tokens at 25 audio tokens per second, with output at roughly 175 text tokens per minute of transcript. At list prices, the blended cost per minute of audio is about $0.005 for gemini-3-5-transcribe and about $0.009 for gemini-3-5-transcribe-live — input at $2 versus $3.50 per million tokens, output at $12 versus $21 per million tokens. Both have a free tier today.

The premium buys the real-time path, not more accuracy: you pay roughly 80% more per minute for the streaming endpoint, and it transcribes at a higher WER. That is the honest frame. The pre-recorded model is both cheaper and more accurate; the streaming model exists because some products cannot wait for the file to finish.

Which endpoint should you build on

• Live captions and subtitles — Ge​mini 3.5 Transcribe Live, and check the no-timestamps limit if you need time-aligned output.

• Voice agents — Ge​mini 3.5 Transcribe Live for intent mid-sentence; plan around the 10-minute cap for long sessions.

• Meetings, interviews, archival — Ge​mini 3.5 Transcribe, for the 2.6% WER, speaker attribution, and word-level timestamps.

• Post-call analytics — Ge​mini 3.5 Transcribe for the finished record; Ge​mini 3.5 Transcribe Live only if the analysis must run during the call.

• Long continuous audio — Ge​mini 3.5 Transcribe, because a 60-minute file beats stitching ten-minute live sessions by hand.

The layer above the transcript

Neither model is something OrcaRouter hosts — we do not route speech-to-text today, and you will call Go​ogle's API directly for either endpoint. What a routing layer does for a voice pipeline is sit above the transcript: the summarizer, the entity extractor, the action-item model that turns a stream into a decision. That layer is where OrcaRouter earns its keep — one API across 200+ models at provider list price with no markup, automatic failover across providers, and a routing DSL that composes several models into a single call. Pick the transcription endpoint by workload, then run the model that reads the transcript behind one key.

Bottom line

Ge​mini 3.5 Transcribe Live and Ge​mini 3.5 Transcribe are the same family and different products. Choose Live when the transcript must exist before the conversation ends and you can live with a 10-minute session, no speaker labels, and no timestamps. Choose Transcribe when the output is a record and accuracy and attribution matter more than speed — it is cheaper, more accurate, and far less constrained. The only wrong answer is assuming one endpoint does both.