Nemotron 3.5 Transcribe vs MAI Transcribe 2: A $100 Hour Against a GPU You Own

One thousand hours of audio costs Microsoft one hundred dollars. That is the whole of the pricing argument for MAI Transcribe 2, which has been in public preview since 3 September 2026 at $0.10 per audio hour — about $1.67 per thousand minutes, a rate that undercuts every managed competitor it names and roughly a third of what OpenAI charges for the same work. Nemotron 3.5 Transcribe, NVIDIA's open checkpoint from 28 October 2026, has no per-minute price at all, because the only way to buy it is to buy the hardware it runs on. At a thousand hours a month, Microsoft's bill is a rounding error on a departmental budget and NVIDIA's is a capital purchase.

Then the awkward part. On the public accuracy board, MAI Transcribe 2 is second in the world at 2.0% AA-WER. Nemotron 3.5 Transcribe is twentieth at 3.4%. The cheaper model is also the more accurate one by a margin larger than the entire gap between ranks nine and twenty. If this comparison were only about words and money, it would be over in a paragraph. It is not, because Microsoft publishes no diarization quality number and NVIDIA publishes several — and that empty column is the reason anyone would still choose the expensive option.

What the board says, and what it cannot say

Artificial Analysis scores both models on the same three-dataset blend, so the 2.0% against 3.4% is at least a like-for-like comparison of the same index, measured by the same operator. That is rarer than it should be in this category, and it deserves to be taken seriously. MAI Transcribe 2 is behind only a real-time preview system at the top of the board; Nemotron 3.5 Transcribe is tied with three systems, one of which is a Gemini Flash-Lite variant in minimal-reasoning mode.

NVIDIA's own launch material reports 2.8% AA-WER, which if it held would still leave it a full eight-tenths behind Microsoft. The live board carries 3.4%. Either way the ordering does not change, so unlike the other matchups in this series, the launch-versus-board discrepancy does not decide anything here. Microsoft is simply ahead on the metric both companies agree to be measured on.

What $0.10 an hour actually buys

MAI Transcribe 2 is a batch transcription model built for long-form audio: meetings, calls, media archives, contact-centre recordings. It covers 60 languages with automatic language identification, bundles speaker diarization, returns word-level timestamps, supports keyword biasing for names and domain terms, and offers two output styles — verbatim, which keeps fillers and false starts, and clean, which strips them. Microsoft reports roughly 411× real-time throughput on Artificial Analysis's speed board, second overall, and its own launch post frames that as ten times faster than GPT Transcribe and five times faster than Gemini 3.5 Transcribe.

The feature list is, on paper, a superset of what most teams need and a near-match for NVIDIA's. Both do diarization, both do word timing, both do keyword biasing, both do verbatim and clean output, both cap around an hour of audio, both are batch-only in the SKU being compared. The differences are at the edges: 60 languages against 40, eight speakers against an unspecified count, and a price that is three orders of magnitude easier to explain to a finance team.

• Price — $0.10 per audio hour, limited-time through the end of 2026, vs no list rate; cost is accelerator time.

• Accuracy — 2.0% AA-WER, rank 2 vs 3.4%, rank 20 (3.4% on the live board; NVIDIA's packet claims 2.8%).

• Throughput — about 411× real time vs up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell with NVFP4 and a draft model, measured on 30-second clips at high concurrency.

• Languages — 60 with automatic identification vs 40 transcribed, 31 translated.

• Licence — proprietary, API-only vs OpenMDW-1.1 with downloadable weights.

• Speakers — diarization bundled, no published DER vs up to 8, attributed, in 13 languages, 18.26% DER across 21 languages on the David AI benchmark.

• Fine-tuning — none offered vs single-GPU adaptation at 0.5–2.3 GPU-hours per task.

• Data locality — audio leaves your network vs self-hostable, including offline.

The column Microsoft leaves empty

Microsoft bundles diarization and publishes no diarization error rate. Not a weak one — none. NVIDIA publishes four separate diarization results: 18.26% DER across 21 languages on the David AI benchmark, lowest of 14 systems; 32.4% concatenated min-permutation WER on VoiceArena's Diarization-Bench, best of 14; 8.32% of reference words dropped on overlapping English speech against a median near 31%; and speaker attribution in 13 languages against Microsoft's unspecified count.

All of those are vendor-reported on a vendor harness, and the honest reading is that they are a claim about capability rather than a proof of quality. But the asymmetry is informative. A company that has a diarization number publishes it, because diarization is the hardest part of a transcript to get right and the part customers complain about. A company that bundles diarization without a number is telling you it does not consider that the selling point. That may mean the feature is adequate and unremarkable. It may mean it is weak. Nothing in Microsoft's launch material distinguishes the two cases, and no independent board measures diarization error rate at this granularity yet.

For a two-person sales call, this does not matter — either model will label two voices correctly. For a six-person board meeting with crosstalk, the difference between 18% DER and an unknown number is the difference between a usable minute-by-minute record and a document nobody trusts. That is the workload where the open checkpoint's price stops looking absurd.

Why anyone buys the expensive one

Four reasons, in descending order of how often they actually appear in procurement decisions.

The first is data locality. MAI Transcribe 2 is an API. The audio leaves your network, lands in a vendor's tenant, and is governed by that vendor's retention terms. Nemotron 3.5 Transcribe runs on hardware you control, including fully offline, which for clinical recordings, legal proceedings, HR investigations and anything under a data-residency regime is not a preference but a requirement. No price per hour competes with a compliance constraint.

The second is permanence. Microsoft's $0.10 rate is explicitly a limited-time offer running through the end of 2026, and no standard price has been disclosed anywhere in the launch material. A team building a product on that number is building on a promotional rate with an unknown expiry. Open weights do not have that failure mode: the checkpoint you downloaded keeps working at the price of electricity, whatever the vendor does next.

The third is adaptation. NVIDIA reports fine-tuning on a single GPU at 0.5 to 2.3 GPU-hours per task, and a medical-speech case study in which adaptation took word error rate from 30.8% to 11.8% across 3,619 clips, with held-out speakers improving from 22.44% to 12.17%. Microsoft offers no fine-tuning path at all. For accented, noisy or jargon-dense audio — exactly the audio where a 2.0% average becomes a 25% reality — the ability to adapt is worth more than two points of leaderboard position. Perit's human-reviewed English set tells the same story from the other direction: it puts MAI Transcribe 2 and GPT Transcribe within a hundredth of a point of each other at 1.76%, which is a reminder that these boards cluster tightly on clean audio and spread out on everything else.

The fourth is that MAI Transcribe 2 is in public preview. Preview means no uptime commitment, no deprecation notice period, and a feature set that can change under you. Nemotron 3.5 Transcribe is one day old and has its own maturity problem — its GGUF tooling is still described as coming soon and its quantised local builds are not shipped — but a checkpoint does not get un-previewed.

The hedge that costs nothing

Most teams do not need to pick. Run the cheap API on the bulk of the queue and route the files that cannot leave the building, or the meetings where speaker attribution has to be right, to the checkpoint. That hybrid is cheaper than either pure strategy and it fails gracefully: if Microsoft's promotional rate expires at an unattractive number, the self-hosted path is already provisioned; if the open model's diarization claim does not survive your audio, the API is already wired in.

What makes that hedge work in practice is keeping the two halves of the pipeline separable. Speech recognition is one contract; the summariser, classifier and router that consume the transcripts are another. OrcaRouter covers the second half with 200+ text and multimodal models behind one key, provider list prices passed through at 0% markup — so a vendor price cut reaches your bill the same day rather than at the next contract renewal — plus automatic failover and a routing DSL that can direct different transcript classes to different models. Neither MAI Transcribe 2 nor Nemotron 3.5 Transcribe is on that catalogue, and we do not serve speech-to-text at all, so the transcription contract stays where you signed it. The point is only that the model downstream of it should not be welded to the same vendor.

The verdict, by volume

Under roughly a few hundred audio hours a month, with no data-residency constraint and no speaker-attribution requirement beyond two-party calls, MAI Transcribe 2 is the better buy on every axis that shows up in a spreadsheet: cheaper, more accurate on the shared index, faster, and available today without provisioning. The promotional pricing is a real risk and worth modelling at two or three times the current rate, but it does not erase a three-order-of-magnitude gap.

Above that volume, or with any of the four constraints above attached, the arithmetic inverts — not because the open model is cheaper per hour, but because the hours stop being purchasable at all. Nemotron 3.5 Transcribe's case rests on speakers, languages, locality, adaptation and permanence, in roughly that order. Its accuracy case rests on a vendor claim the public board does not currently support, and anyone repeating "first among open weights" as if it settled the comparison is quoting a slide rather than a measurement.