Grok Voice Transcribe 2.0 vs MAI Transcribe 2: Two Lanes, Not Two Rivals

These two models were released fifteen days apart and priced to the cent the same, and they will almost never be in the same procurement decision. Grok Voice Transcribe 2.0, shipped by SpaceXAI on September 18, 2026, is the model you call when audio is still arriving — it streams over a WebSocket, emits interim results roughly every 500 ms, and tops Artificial Analysis's AA-WER Streaming board at 2.7% word error rate for final transcripts and 3.4% for first partials. MAI-Transcribe-2, released by Microsoft AI on September 3, 2026, is the model you call when the audio is already a file — it is batch-only, and on Artificial Analysis's non-streaming board it is the most accurate managed model measured at 2.04% AA-WER, ahead of Grok Voice Transcribe 2.0's 2.29% and roughly 2× faster in throughput. Both list at $0.10 per audio hour, about $1.67 per 1,000 minutes. If your requirement is real-time, there is no comparison to make. If your requirement is a file, there is.

The line neither vendor will draw for you

Streaming support is not a feature toggle. Grok Voice Transcribe 2.0 serves the same model over REST for recorded files and over `wss://api.x.ai/v1/stt` for live audio, so the accuracy you measure on your archive is the accuracy you get on a live call. MAI-Transcribe-2 has no streaming SKU at all: Microsoft's launch materials position it explicitly for long-form and batch audio, and while the model card lists real-time captioning among its application scenarios, the route to that is segmenting audio and feeding each segment through the batch API — which is a pipeline you build and operate, not a latency guarantee you buy. Microsoft's headline throughput of roughly 411× real time is a processing-speed figure, not a response-time one, and the two get conflated constantly in coverage of this release.

The practical consequence: if you are building turn-taking voice agents, live captioning, or anything where a human or a model reacts to speech in progress, MAI-Transcribe-2 is not a candidate no matter how good its accuracy is. If you are transcribing meetings after the fact, contact-centre recordings overnight, media archives, or compliance backlogs, streaming is dead weight and the batch numbers are the whole story.

Where MAI-Transcribe-2 is genuinely ahead

Accuracy on files, first. The 2.04%-versus-2.29% gap is measured on the same independent harness — Artificial Analysis's AA-WER v2, 50% AA-AgentTalk and 25% each VoxPopuli and Earnings22 — which makes it the cleanest fact in this comparison. It is a 0.25-point difference, roughly two and a half fewer errors per thousand words. Whether that matters depends on what you do with the transcript; for a searchable archive it does not, and for a legal or medical record it might.

Throughput, second, and by more than the accuracy gap: 333× real time against Grok Voice Transcribe 2.0's 162×. Both numbers are Artificial Analysis measurements of input audio seconds transcribed per second, not vendor claims. At that rate a thousand-hour archive finishes in about three hours of compute on the Microsoft model and about six on the SpaceXAI one. For a one-off migration that is irrelevant; for a pipeline that ingests continuously, the halved wall-clock is a real capacity difference.

Language coverage, third. Microsoft specifies 60 languages with automatic detection, code-switching within a recording, and an average 5.2% word error rate across them on FLEURS — a vendor-reported number, not independently reproduced, but at least it describes the multilingual case across a stated language list. SpaceXAI documents dozens of languages with mid-recording switching and written-form formatting across 25. If your audio is outside the well-covered English-and-major-European set, the FLEURS figure is more informative than a leaderboard that is English-weighted and agent-talk-heavy.

Two more differences that matter operationally. MAI-Transcribe-2 offers configurable transcription styles — verbatim, preserving disfluencies, versus clean, which strips them — where Grok Voice Transcribe 2.0 handles the same problem with a filler-word-removal flag. And Microsoft's model is in public preview on Microsoft Foundry, which means no SLA and a standing recommendation against production use until it goes GA; SpaceXAI's model is generally available with published rate limits of 10 requests per second and 100 concurrent streaming sessions per team.

Where Grok Voice Transcribe 2.0 is genuinely ahead

• Real-time streaming — a WebSocket endpoint with interim results every ~500 ms, versus batch-only with no real-time SKU announced

• First partial transcript accuracy — 3.4% WER at 0.49s on AA-WER Streaming, a category MAI-Transcribe-2 does not compete in

• Batch accuracy on agent-talk audio — 2.29% overall, but on the AA-AgentTalk subset of the non-streaming board the ordering is tighter than the headline suggests, and on the streaming board's agent-heavy mix Grok leads outright

• Voice-agent plumbing — Smart Turn end-of-turn detection and filler-word removal ship in the model, where MAI-Transcribe-2 leaves turn logic to the caller

• Multichannel audio — up to 8 independent channels transcribed in one pass, which matters for stereo or multi-leg call recordings

• Word-level confidence — timestamps carry a confidence score per word, so low-confidence spans can be flagged rather than trusted equally

• Availability terms — generally available with published concurrency limits, against a public preview with no SLA

• Price durability — $0.10 per hour is the standing rate for both batch and the base for the $0.20 streaming tier, where Microsoft's identical $0.10 is a limited-time launch offer with no announced standard rate for 2027

That last point is the one most comparisons skip. The two models cost the same today, and one of those prices has an expiry date the other does not. Microsoft has not published a post-promotion rate, and coverage disagrees on whether the offer runs through the end of 2026 or has no stated end at all. If you are modelling a three-year transcription budget on $1.67 per 1,000 minutes from Microsoft, you are modelling a number the vendor has not committed to.

The overlap is real, and it is most of the feature list

Set the streaming question aside and the two products converge hard. Both do speaker diarization. Both return word-level timestamps. Both handle inverse text normalization — numbers, dates, currencies, contact details rendered in written form. Both accept domain terminology, Grok Voice Transcribe 2.0 as up to 100 key terms per request and MAI-Transcribe-2 as domain terminology and abbreviation lists. Both detect language automatically and follow a speaker switching languages mid-recording. Both handle 8 kHz telephony audio, which is the case most transcription models quietly fail on and the one SpaceXAI tuned hardest against — its internal telephony set went from 10.6% to 7.1% WER between versions, a company-reported figure on company audio.

Both also carry input limits that shape architectures rather than just budgets. Grok Voice Transcribe 2.0 accepts files up to 500 MB across 12 audio formats with sample rates from 8 kHz to 48 kHz. MAI-Transcribe-2's limits are set by the Foundry path rather than the model, and teams should read Microsoft's own documentation for the current ceiling rather than assume parity with a dedicated STT service.

What that convergence means is that the decision reduces to three questions in order: does it need to be live, how many languages do you actually have, and how much does the accuracy gap cost you. The first question is binary and eliminates one option outright. The second is usually answerable from a sample of your own audio. The third is small enough that for most file-based workloads it will not decide anything — the 0.25-point gap is real, but so is the fact that both models are within half a point of the best number anyone has measured.

What to do with this

Treat them as a two-lane stack rather than a head-to-head. A contact centre that runs live agent assistance and an overnight compliance archive is not choosing between Grok Voice Transcribe 2.0 and MAI-Transcribe-2 — it is running both, and the only question is whether that costs it two vendor relationships, two sets of credentials, two invoices and two rate limits. Neither model is in OrcaRouter's catalogue today, so the transcription calls themselves go to each vendor's own API. What one OrcaRouter key does cover is everything downstream: the summarizer, the classifier, the agent that reasons over the transcript, and the eval job that compares the two vendors' output on the same recording. That last one is the reason to care — you cannot decide between these models from a leaderboard, because the leaderboard is eight hours of English agent talk and your audio is not.

Watch two things. First, whether Microsoft attaches a standard price to MAI-Transcribe-2 when the launch offer ends; a permanent $1.67 per 1,000 minutes would make it the default batch choice on cost alone, and a reversion toward MAI-Transcribe-1's $0.36 per hour would make this a much shorter conversation. Second, whether Microsoft ships a streaming SKU. The model card already names real-time captioning as a target scenario, and the only thing standing between MAI-Transcribe-2 and the streaming lane is an endpoint — if that lands, these two stop being lanes and start being rivals.