Nemotron 3.5 Transcribe vs Grok Voice Transcribe 2.0: Open Weights on Your GPU, or $0.20 an Hour

Grok Voice Transcribe 2.0 bills $0.20 for every hour of audio you stream through it, and $0.10 for every hour you batch. Nemotron 3.5 Transcribe bills nothing per hour and asks you to own a GPU that can hold about 34 billion parameters. Read only that sentence and the comparison looks like a price war between a meter and a machine, with the meter obviously losing. It does not lose: on the one benchmark board where both models appear, the metered one is more accurate, and the open one is not close enough to argue the point.

The two shipped six weeks apart — Grok Voice Transcribe 2.0 on 2026-09-18, Nemotron 3.5 Transcribe on 2026-10-28 — and they were built for different jobs. One is a hosted API with a WebSocket that returns words while a person is still talking. The other is a checkpoint you download, load, and point at a finished recording. Latency, language count, speaker labels, cost structure and even the legal terms all follow from that single difference, which is why the interesting question is not which one transcribes better.

What each one is, in six lines

The shape contrast is stark enough that a spec list reads like a category error, which is the point.

• Deployment — a hosted REST and WebSocket API on xAI's own infrastructure versus open weights under OpenMDW-1.1 that you serve yourself on vLLM, SGLang-Omni or NeMo Speech.

• Size — xAI publishes no parameter count versus 34B total with roughly 4.3B active, including the audio encoder, sitting on a Nemotron 3.5 Lightning 30B-A3B Mamba and Transformer mixture-of-experts backbone.

• Audio in — files up to 500MB in twelve container formats plus three raw encodings, with 2-8 channels via multichannel mode, versus recordings up to one hour at 16kHz mono PCM inside a 65,536-token context.

• Realtime path — streaming with interim results roughly every 500ms, endpointing defaulting to 400ms, smart turn detection, and a separate batch endpoint, versus no streaming at all; the model reads a finished file and returns a transcript.

• Language coverage — 37 languages named in xAI's documentation and described there as 38+, versus 40 languages transcribed and 31 translation pairs to or from English on the NVIDIA card.

• Speaker labels — a diarize flag that adds a speaker integer to every word, available on batch and streaming, versus a fused diarization encoder that produces speaker-attributed transcripts in one pass for up to 8 speakers in 13 languages.

• Price — $0.10 per audio hour in batch and $0.20 per audio hour streaming, versus no published per-hour rate, because NVIDIA sells weights and you supply the compute.

The one board where they meet, and it does not break the way you expect

Artificial Analysis runs a speech-to-text leaderboard with two lanes: a batch lane for finished files and a streaming lane for models that emit partials. Nemotron 3.5 Transcribe only qualifies for the first. There, Grok Voice Transcribe 2.0 sits at 2.3% AA-WER, fifth of roughly sixty models, while NVIDIA's model is quoted at 2.80% in the launch material and reads 3.4% on the live table in the days after release. Call it 2.8% or 3.4%; either way the hosted model wins the only scale both of them share, by between half a point and a full point.

That result is less damaging than it first looks, because the two models are not on the board for the same reason. NVIDIA's claim is that Nemotron 3.5 Transcribe is the best open-weight speech model, and on that narrower board it is: the comparison it wins is against other downloadable checkpoints, where 2.8% is the number to beat. Grok Voice Transcribe 2.0 is proprietary and competes in a different lane. Quoting the two against each other produces a headline — "hosted model beats open model by 0.5 points" — that measures a benchmark mix, not a product decision.

The streaming lane is where the asymmetry becomes structural rather than statistical. Grok Voice Transcribe 2.0 holds the top streaming slot at 2.7% final WER, emitted 0.49 seconds after the end of speech, and 3.4% on its first partial at the same latency, first of about thirty-two streaming models. Nemotron 3.5 Transcribe has no entry, because it cannot stream. It is a file-in, transcript-out model, and no chunked streaming path is documented for it. Anyone comparing the two on accuracy while planning a live-caption deployment is comparing a model they cannot use.

Both boards carry a warning label worth repeating. AA-WER moves by nearly a full point for the same model across different serving endpoints, so the 0.5-point gap above is roughly the width of the measurement noise — real in direction, weak as a verdict. And xAI's own 6.8% multilingual short-phrase figure, down from 20.6% in version 1.0, is drawn from an internal set built out of production traffic, not a released corpus. It is a vendor claim about a vendor set, and it does not appear on Artificial Analysis at all.

Streaming is the line the open model cannot cross

For a call-centre agent, a live caption, a meeting assistant or a voice bot, the decision is over before accuracy is considered. Grok Voice Transcribe 2.0 accepts raw binary audio frames over a WebSocket — pcm, mulaw, alaw or opus — returns interim results about twice a second, and provides turn-end signals through endpointing and smart turn detection. Those are the mechanics that make a transcript feel live, and they are the reason Atlassian's Loom moved its video library onto the model at launch.

Nemotron 3.5 Transcribe has the opposite shape by design. It reads up to an hour of audio in one pass and returns a complete transcript, with word timestamps produced by a lightweight CTC adapter after generation rather than during it. That design buys accuracy on long, messy recordings and costs latency on everything short and interactive. NVIDIA has a separate, much smaller streaming speech recognizer in its line — Nemotron Speech Streaming EN 0.6B — but it is a different model with different coverage, not a streaming mode of this checkpoint. Buying the 30B model expecting to bolt streaming on later is buying the wrong thing twice.

Forty languages against thirty-seven, and a count that moved

On paper the gap is three languages. In practice it depends which page you read. xAI's speech-to-text documentation lists 37 named languages — from Arabic and Bosnian through Urdu and Vietnamese, with regional variants for Spanish, Portuguese and Arabic — and describes the total as 38+. Coverage published around the September launch said 25. The documentation is the more current source and the more specific one, but a language count that changes by thirteen between a launch post and a docs page is a count to verify against your own audio rather than adopt.

NVIDIA's card is more granular and, for once, more conservative: 40 languages with a per-language checkmark table, and 13 of those flagged for multispeaker transcription. The ones that matter for speaker attribution are a narrower set than the 40 — Arabic, Mandarin, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Spanish and Thai. If your archive is in Polish or Turkish, Nemotron 3.5 Transcribe will transcribe it well and will not tell you who said what, which is the capability you were probably buying.

Translation is a Nemotron-only column here. The model translates to and from English across 31 pairs in the same pass that transcribes, which Grok Voice Transcribe 2.0 does not offer; xAI's API returns text in the language spoken. If the job is bilingual meeting minutes, that is a feature no prompt engineering on the Grok side will reproduce.

Speaker labels: a fused encoder against a boolean

This is the one dimension where the two are closer than the marketing suggests, and where the honest answer is that both do it.

• Nemotron 3.5 Transcribe — a frozen Nemotron 3 Diarization model feeds speaker-activity representations into the ASR encoder, so the language model sees who is speaking as part of its input. Output carries speaker tags and raw segments, up to 8 speakers, and word-level timestamps. It is one pass with no second service in the middle.

• Grok Voice Transcribe 2.0 — set diarize=true and every word comes back with an integer speaker field, on batch and streaming alike. Multichannel mode goes further: up to 8 channels transcribed independently and returned with a channel index, which is exact attribution when your source is a multichannel recording rather than a mixed one.

The practical difference is coverage, not existence. Grok's flag works across the languages its API supports; NVIDIA's fused path is trained for 13. And multichannel is a real advantage in the rooms where it applies — a boardroom with one microphone per seat does not need a diarizer to guess. A mixed mono file from a conference floor does, and there the quality gap between a fused encoder and a hosted flag is something only your own audio can settle.

The arithmetic at a thousand hours a month

Price is where the two models genuinely diverge, and it is worth doing the sum rather than trading anecdotes about GPU bills.

At 1,000 audio hours per month, Grok Voice Transcribe 2.0 costs $100 in batch mode or $200 streaming — $1,200 or $2,400 a year, with no hardware, no maintenance window and no cold start. Nemotron 3.5 Transcribe costs whatever the GPU costs. NVIDIA's research page claims the NVFP4 checkpoint with the DFlash2 draft model transcribes 2,238 seconds of audio per second on a single RTX PRO 6000 Blackwell, which even discounted by an order of magnitude puts 1,000 hours of audio inside a few GPU-hours per month. Compute is not the constraint; ownership is.

• Break-even against the streaming rate — an $8,000 workstation card pays for itself after 40 months of 1,000 streaming hours, or 80 months against the batch rate.

• At 10,000 hours a month — the same card breaks even in four months against streaming, and the per-hour rate falls to about 2 cents.

• Rented cloud GPU — an H100 at roughly $2 an hour beats both models' list prices the moment the vendor's throughput claim holds at even a fraction of its stated figure; the fixed cost becomes engineering time, storage and the fact that you now run a model server.

The break-even point is therefore not about transcription accuracy at all. It is about volume and about whether the audio is allowed to leave your network. Under a few thousand hours a month, the hosted rate is cheaper than the engineering. Above it, or behind a compliance wall, the weights win and the comparison flips.

Where OrcaRouter fits, and where it does not

Neither of these transcribers is in OrcaRouter's catalogue today; the models we serve are text and vision models, and nothing here should be read as an endpoint for either one. What a transcription pipeline does generate is a large pile of text that then needs summarising, entity extraction, redaction, classification and search. Those calls are where a single OpenAI-compatible key across 200+ models earns its keep: provider list price passed through at 0% markup, so a vendor price cut lands the same day, automatic failover when one upstream degrades, and a routing DSL that can send cheap classification work to a small model while holding the expensive summarisation on a frontier one. If you self-host Nemotron 3.5 Transcribe, the GPU is your transcriber; routing is what happens to the transcript afterwards.

Which one to pick

The decision is about the clock on the audio, not the model card.

• Live audio, voice agents, captions, call routing — Grok Voice Transcribe 2.0, without much debate. Streaming is native, latency is under half a second to a final transcript, and the batch endpoint is the same model for the overnight backfill.

• Finished recordings with multiple speakers, in languages outside English — Nemotron 3.5 Transcribe, provided you have the GPU and the 13-language speaker path covers your audio. This is the case the model was built for, and where 40 languages and one-pass diarisation beat a 0.5-point accuracy gap.

• Bilingual archives — Nemotron 3.5 Transcribe, because the translation pairs ship inside the same pass and Grok does not translate.

• Small volume, no GPU, English only — Grok Voice Transcribe 2.0 in batch mode at $0.10 an hour. A thousand hours costs a hundred dollars; no amount of open-weight licence freedom makes up for buying a Blackwell card to save that.

The tempting fourth answer — deploy Nemotron now and add streaming later — does not exist. The model NVIDIA shipped is deliberately offline, and the streaming line it sells alongside is a 0.6B English recognizer, not a mode of this one. Pick the shape you need first; the leaderboard argument can wait, because half a point of AA-WER is well inside the range the same model moves when it changes hosts.