Grok Voice Transcribe 2.0 Takes the #1 Streaming Accuracy Slot at an Unchanged Price

Grok Voice Transcribe 2.0 is the speech-to-text model SpaceXAI shipped on September 18, 2026, and the single number that matters about it is not the one the launch post leads with. It is this: the first partial transcript — the text a voice agent actually gets to act on while the caller is still being talked over — went from 18.3% word error rate in Grok Voice Transcribe 1.0 to 3.4%. That is the measurement on Artificial Analysis's AA-WER Streaming board, where the new model now sits first on both the Final Transcript ranking (2.7% WER, 0.49 seconds after end of speech) and the First Partial Transcript ranking (3.4%, 0.49 seconds). Final-transcript accuracy improved too, from 3.9% to 2.7%, which is real but modest. The partial-transcript jump is the one that changes what you can build.

What actually changed between 1.0 and 2.0

The two versions are the same product in the same API, and SpaceXAI made no changes to the interface: Grok Voice Transcribe 2.0 is selected by passing grok-voice-transcribe-2.0 to /v1/stt instead of grok-voice-transcribe-1.0, or by choosing it when the WebSocket connection is created for streaming. Existing integrations that omit the model parameter keep getting 1.0 until SpaceXAI flips the default, which the company says will happen in the coming weeks alongside deprecation of the older version. Until then 2.0 is opt-in and 1.0 can be pinned deliberately, which is the right shape for a change like this: you can A/B it on your own audio before it becomes your production default whether you asked or not.

The Artificial Analysis numbers, read side by side, tell a more specific story than "twice as accurate":

• Final transcript accuracy — Grok Voice Transcribe 2.0 at 2.7% WER vs Grok Voice Transcribe 1.0 at 3.9% on AA-WER Streaming, both measured after end of speech

• First partial transcript accuracy — 3.4% WER vs 18.3%, the largest single gap between the two versions

• Time to final transcript — 0.49s vs 0.37s, so the new model is slower to commit than the one it replaces

• Time to first partial — 0.49s vs 0.25s, likewise slower

• Batch (non-streaming) accuracy — 2.3% AA-WER on Artificial Analysis's non-streaming board, against 4.07% for Whisper Large v3 and 3.31% for GPT Transcribe

• Batch throughput — roughly 162× real time on the same board, behind MAI-Transcribe-2's 333× and ahead of Gemini 3.5 Transcribe's 88×

That fourth and fifth line are the honest caveat in this release. SpaceXAI bought accuracy with latency. The new model is about a third of a second slower to return its first partial and its final transcript than 1.0 was. On a board where Deepgram Flux returns a partial in 0.02 seconds and Soniox v5 in 0.05, Grok Voice Transcribe 2.0 is not competing on speed at all — it is competing on being right, and it wins that trade by a wide margin. Whether that is the right trade depends entirely on whether your application is a live voice agent that needs to start speaking back, or a transcription pipeline where half a second is invisible.

The "twice as accurate" claim, checked

SpaceXAI's headline is that 2.0 is twice as accurate as 1.0 at the same price, and its own evaluation table backs that up on the sets it chose. On 8 kHz English customer-support calls, word error rate fell from 10.6% to 7.1%. On conversations with Grok, from 8.7% to 3.3%. On spoken credentials — account codes, phone numbers, email addresses — from 7.2% to 3.2%. On short commands across 19 languages, from 20.6% to 6.8%, a roughly two-thirds reduction in errors. Those four sets are drawn from SpaceXAI's production traffic and the comparisons are SpaceXAI's own; nobody outside the company has reproduced them. Treat the table as a map of where the model was tuned, not as a general accuracy guarantee.

The Artificial Analysis result is the part that is not vendor-reported: an independent harness run against roughly eight hours of audio weighted 50% to the private AA-AgentTalk set and 25% each to VoxPopuli and Earnings22. It supports a narrower and more useful claim than "twice as accurate" — that on that specific mix, this model is the most accurate streaming transcriber measured. One wrinkle worth naming: SpaceXAI's post describes a first place "among 32 streaming models," while the leaderboard page itself plots 27 of 33 models. The rank is real; the field size in the marketing copy is the company's count, not the board's.

It is also worth being precise about what a first place on AA-WER Streaming does and does not cover. The board is English-weighted and agent-talk-heavy. It does not measure your accent, your codec, your domain vocabulary, or your eight-channel call-centre audio. A model that tops it can still lose badly on your recordings, and that is exactly why the opt-in window matters.

Pricing is unchanged, and that is the second-biggest fact here

Grok Voice Transcribe 2.0 costs what 1.0 cost: $0.10 per audio hour for batch and recorded files over the REST endpoint, $0.20 per audio hour for streaming over the WebSocket. On Artificial Analysis's price board the streaming SKU lands at $3.33 per 1,000 minutes and the batch SKU at $1.67 — the same batch rate Microsoft is charging for MAI-Transcribe-2, and roughly a third of what ElevenLabs Scribe v2 Realtime costs per streaming minute.

Diarization, word-level timestamps with confidence scores, and key-term biasing are all included at that rate rather than metered separately, which is not universal in this market. Gemini 3.5 Transcribe Live bills per token on both audio input and text output, so your cost moves with how much the model says, not just how long the audio ran. A flat per-hour rate is easier to forecast and easier to compare across vendors — and because OrcaRouter passes provider list prices through with 0% markup, a vendor-side change to that rate would show up on our side the same day rather than after a repricing cycle.

What the API actually gives you

The capability list is broad and mostly inherited rather than new, which is worth knowing if you are evaluating the upgrade on features alone:

• Batch and streaming in one model — REST for recorded files up to 500 MB, WebSocket at wss://api.x.ai/v1/stt with interim results emitted roughly every 500 ms

• Speaker diarization included, plus multichannel transcription of up to 8 independent channels

• Word-level timestamps carrying confidence scores, so you can threshold low-confidence spans instead of trusting the whole transcript equally

• Key-term biasing of up to 100 domain terms per request, each up to 50 characters

• Inverse text normalization for numbers, dates, currencies, phone numbers and email addresses, with written-form formatting supported across 25 languages

• Filler-word removal and Smart Turn end-of-turn detection aimed squarely at voice agents

• Automatic language detection with mid-recording language switching in a single pass, across 12 audio formats and sample rates from 8 kHz to 48 kHz

• Service limits of 10 requests per second on both REST and streaming, and 100 concurrent streaming sessions per team, hosted in us-east-1

The one production note that is not on the feature list: the service runs in a single region. If your audio originates outside the United States, the network leg is now part of your latency budget, and AA-WER Streaming explicitly folds network delay into its timing. SpaceXAI offers zero-data-retention terms and cites SOC 2 Type II, BAAs and EU data residency options, so the compliance story is more developed than the geographic one.

Who should move, and what to watch

If you are already on Grok Voice Transcribe 1.0 and your application acts on partial transcripts — turn detection, barge-in, live agent responses — the partial-accuracy gap from 18.3% to 3.4% is large enough that testing 2.0 is not optional. That number going from "usually wrong" to "usually right" is the difference between a partial transcript you can route on and one you can only display. If your application waits for final transcripts on recorded files, the improvement is real but smaller, and the added 0.12 seconds of commit latency is the cost of it.

If you are on a different vendor entirely, the reason to look at this release is not the leaderboard rank — it is that a frontier lab just improved its transcription model by a wide margin without raising the price, which is what a market looks like when accuracy stops being the thing you charge for. The upgrade path is also the cheapest kind: one parameter, no code change, no new contract, and a pin available if it goes wrong.

What to watch next is the default flip. When SpaceXAI makes 2.0 the default and starts deprecating 1.0, every integration that never passed a model parameter will move onto the new model at once, latency change included. Teams that depend on the old 0.25-second partial timing should pin now rather than discover the change in production. The second thing to watch is independent reproduction: the Artificial Analysis entry is a real measurement, but it is one board on one audio mix, and no third party has yet published a like-for-like 1.0-versus-2.0 run on production audio.