Grok Voice Transcribe 2.0 vs Gemini 3.5 Transcribe: The Streaming Lane Belongs to One of Them

Grok Voice Transcribe 2.0 and Gemini 3.5 Transcribe are the two newest frontier speech-to-text models from rival labs, and the honest way to compare them is to admit up front that they are not built for the same job. SpaceXAI's Grok Voice Transcribe 2.0, released September 18, 2026, is a streaming-first model with a batch mode attached: it tops the AA-WER Streaming board at 2.7% word error rate for final transcripts and 3.4% for first partials, both arriving 0.49 seconds after end of speech. Gemini 3.5 Transcribe, released August 26, 2026, is a batch-first model with a streaming SKU attached, and the two halves of it behave very differently — the pre-recorded path measures 2.6% AA-WER on Artificial Analysis's non-streaming board, while the separate gemini-3.5-transcribe-live endpoint measures 4.0% on the streaming board at 0.40 seconds. Put those four numbers next to each other and the shape of this matchup is obvious: they are close on accuracy where Gemini 3.5 Transcribe is strong, and they are not close where Grok Voice Transcribe 2.0 is.

The only lane they both fight in

Both models can transcribe live audio, so streaming is the one arena where a straight comparison means something. There, Grok Voice Transcribe 2.0 leads Gemini 3.5 Transcribe Live by 1.3 points on final-transcript accuracy — 2.7% versus 4.0% WER on the same harness, the same roughly eight hours of audio, and the same 50/25/25 weighting across AA-AgentTalk, VoxPopuli and Earnings22. That is not a rounding difference; at those error rates it is roughly one extra wrong word in every seventy-five transcribed by the Google model. On first partial transcripts the gap is wider still, 3.4% against 5.8%, and partials are the transcripts a live voice agent has to act on before the speaker has finished.

Latency cuts the other way, and this is the part of the comparison that gets lost. Gemini 3.5 Transcribe Live returns its final transcript 0.40 seconds after end of speech against Grok's 0.49, and its first partial in 0.25 seconds against Grok's 0.49 — roughly twice as fast to the first usable word. Neither model is competing with the genuinely fast end of this board, where Deepgram Flux posts 0.02 seconds and Soniox v5 0.05, but between these two the trade is explicit: Google is quicker to speak, SpaceXAI is more likely to be right when it does. For a turn-taking agent that must not talk over a caller, 0.24 seconds of extra partial latency is a real cost that the accuracy win has to justify.

Where Gemini 3.5 Transcribe actually wins

Language coverage is the clearest one, and it is not close. Google's model handles 85+ languages; Grok Voice Transcribe 2.0 supports dozens, with written-form formatting — the inverse text normalization that turns spoken numbers, dates, currencies and email addresses into their written shapes — documented across 25 languages. If your audio is Portuguese, Vietnamese, or a code-switched mix of two languages inside one sentence, the question of which model is more accurate on the Artificial Analysis board is largely academic, because that board is English-weighted and agent-talk-heavy. Google reports 5.50% streaming and 5.04% non-streaming WER on FLEURS across its own multilingual suite, figures that are vendor-reported and not independently reproduced but at least describe the multilingual case at all.

The second win is the free tier. Gemini 3.5 Transcribe carries free input and output tokens on Google's API, with the caveat that free-tier content is used to improve Google products while paid-tier content is not. For a prototype, a side project, or an internal tool processing audio you would not otherwise pay to transcribe, that is a real option no per-hour vendor can match. Grok Voice Transcribe 2.0 is paid from the first audio hour.

And the third is that Gemini 3.5 Transcribe is a general multimodal model with a transcription mode, not a purpose-built ASR service. It can be prompted, it can take text as context, and it lives in the same API surface as everything else Google ships. If transcription is one step in a pipeline that also needs reasoning over the transcript, keeping both in one model call is worth something that a per-word error rate does not capture.

The price math, per thousand minutes

Artificial Analysis normalizes both models to cost per 1,000 minutes of audio, which is the only way to compare a flat hourly rate against token billing:

• Batch, pre-recorded — Grok Voice Transcribe 2.0 at $1.67 per 1,000 minutes ($0.10/hour) vs Gemini 3.5 Transcribe at $5.00 per 1,000 minutes ($0.30/hour, from $2.00 per 1M input and $12.00 per 1M output tokens)

• Streaming — Grok Voice Transcribe 2.0 at $3.33 per 1,000 minutes ($0.20/hour) vs Gemini 3.5 Transcribe Live at roughly $5.40 per 1,000 minutes, blended from $3.50 per 1M audio-input and $21.00 per 1M text-output tokens

• Batch throughput — Grok Voice Transcribe 2.0 at roughly 162× real time vs Gemini 3.5 Transcribe at roughly 88× on the same board, so the cheaper model is also the faster one for file work

• Live session cap — Grok Voice Transcribe 2.0 streams over a WebSocket with no published session ceiling beyond 100 concurrent sessions per team vs Gemini 3.5 Transcribe Live capped at 10 minutes per session

That last line is the operational detail that decides more architectures than the accuracy numbers do. A 10-minute ceiling on a live session means long calls, long meetings, and long streams must be chopped and re-established, and every re-establishment is a seam where context is lost. Grok's streaming endpoint carries a file-size ceiling of 500 MB on the batch path and a per-team concurrency limit of 100 sessions on the streaming path, which is a different kind of constraint — throughput rather than duration.

Token billing is worth understanding before you commit to the Google side, because it makes your bill a function of two variables instead of one. Gemini charges for audio input and for text output separately, so a model that produces verbose formatting or repeats itself costs more than one that does not, and a language with longer words costs more than one with shorter ones. Grok's flat per-hour rate does not move with any of that. For high-volume workloads the flat rate is easier to forecast, and because OrcaRouter passes provider list prices through with 0% markup, a change on either vendor's side reaches your bill the day it happens rather than at the next contract renewal.

Feature shapes: what each one refuses to do

The capability lists diverge more than the accuracy numbers suggest, and the gaps are in places that matter to specific applications:

• Speaker diarization — included at no extra cost on Grok Voice Transcribe 2.0; not supported on Gemini 3.5 Transcribe Live, so speaker-attributed live transcripts are not available on that endpoint at all

• Word-level timestamps — Grok Voice Transcribe 2.0 returns them with per-word confidence scores; Gemini 3.5 Transcribe Live returns none, which rules out confidence thresholding and precise caption alignment

• Multichannel audio — Grok Voice Transcribe 2.0 transcribes up to 8 independent channels in one pass; Gemini 3.5 Transcribe Live has no equivalent, so multi-channel call recordings need per-channel sessions

• Key-term biasing — Grok Voice Transcribe 2.0 accepts up to 100 domain terms per request; Gemini 3.5 Transcribe accepts custom vocabulary and optional language hints

• Voice-agent plumbing — Grok Voice Transcribe 2.0 ships filler-word removal and Smart Turn end-of-turn detection; Gemini 3.5 Transcribe Live covers the streaming case but leaves turn logic to the application

• Input handling — Grok Voice Transcribe 2.0 takes direct file upload up to 500 MB across 12 audio formats; Gemini 3.5 Transcribe's pre-recorded path expects a public HTTPS URL with a 15-minute and 300 MB ceiling, and cannot combine its Smart formatting mode with word timestamps or speaker labels

The pattern in that list is that SpaceXAI built a transcription product and Google built a transcription capability. Diarization, timestamps, channel splitting and turn detection are the features you need when transcription is the whole application; promptability, language breadth and one-API-for-everything are what you want when transcription is a step inside a larger one. Neither list is better in the abstract, and a team that picks on accuracy alone will end up missing whichever set it did not need.

Picking between them, and what changes the answer

Reach for Grok Voice Transcribe 2.0 when the transcript has to be right while the audio is still playing: live agent turn-taking, real-time captioning that cannot afford to be wrong, contact-centre streams where speaker attribution and channel separation are part of the requirement, or any batch pipeline where $1.67 per 1,000 minutes and 162× throughput both matter. Reach for Gemini 3.5 Transcribe when the audio is multilingual in a language outside Grok's documented set, when the transcript feeds a model that also has to reason about it, when you want a free tier to prototype against, or when the batch path's 2.6% AA-WER is simply enough for what you are doing and the extra language coverage is worth more than the price difference.

What most teams actually do here is not pick. Both models are reachable through one OrcaRouter API key alongside 200+ other models, which means you can route live agent traffic to Grok Voice Transcribe 2.0 and long multilingual archives to Gemini 3.5 Transcribe without maintaining two vendor contracts, two billing relationships, and two sets of credentials. That is also how you settle the accuracy question for your own audio: run both on the same recordings, compare the transcripts you actually got, and let the routing DSL send the traffic where it belongs. The leaderboard tells you which model wins on eight hours of someone else's English agent talk; only your own audio tells you which one wins on yours.

The open question is what happens to that streaming gap when Google next revises the Live endpoint. Gemini 3.5 Transcribe Live launched in public preview and its 4.0% figure was measured roughly a day after it became callable — a strong early signal, but one that has had less time to settle than the Grok number it now sits behind. If Google closes the 1.3-point streaming gap while keeping the 0.25-second partial latency, the trade stops being a trade and Grok's advantage narrows to price and features. Until then the split is clean: the fastest partials are Google's, the most accurate ones are SpaceXAI's, and no model on the board is both.