Nemotron 3.5 Transcribe vs GPT Transcribe: 0.1 Points Apart, One Hands You the Checkpoint

Ask GPT Transcribe for word-level timestamps and the API will not give them to you. Not badly, not approximately — the model has no timing head, OpenAI's documentation points you at whisper-1 for that job, and the same is true of speaker labels, which live in a separate diarization model. Now ask Nemotron 3.5 Transcribe, which NVIDIA released as open weights on 28 October 2026, and both come back from a single pass: speaker-attributed word timing from a CTC adapter that aligns the finished transcript to the audio after decoding. The two models are within a tenth of a point of each other on the public accuracy board. On the things you do with a transcript, they are not in the same product category.

That is the whole decision, and it is worth walking through slowly, because the launch coverage has mostly reduced this pairing to a WER race that neither side actually wins by a meaningful margin.

The accuracy gap is noise; the feature gap is not

On Artificial Analysis's non-streaming index, GPT Transcribe — OpenAI's batch transcription model, shipped 28 July 2026 alongside a live-streaming sibling — is listed at 3.3%, tied at rank eighteen with a small challenger. Nemotron 3.5 Transcribe is listed at 3.4%, tied at rank twenty. One tenth of a point, on an index whose own entries for the same model can differ by nearly a full point depending on which endpoint served them. Nobody should choose between two speech systems on that.

NVIDIA's launch material, meanwhile, reports 2.8% AA-WER and a tie for first among open-weight models. The live board read on 29 October carries 3.4%. Treat the 2.8% as a submission figure and the 3.4% as the current measurement; the difference is larger than the gap between Nemotron and GPT Transcribe, which tells you how much weight a single leaderboard delta deserves.

• Parameters — about 30B total and roughly 4B active per token in a hybrid Mamba–Transformer MoE vs undisclosed, non-generative.

• Licence — OpenMDW-1.1, commercial use, weights downloadable vs API-only.

• Price — no list rate; GPU time, with an NVFP4 build claimed at up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell vs $0.0045 per minute, about $0.27 per audio hour.

• Speaker labels — up to 8 speakers in 13 languages, built in vs none; a separate diarization model is required.

• Word timing — post-decode CTC alignment, speaker-attributed vs not offered; whisper-1 is the documented alternative.

• Languages — 40 transcribed, 31 translated to or from English vs multilingual with auto-detection and language hints, no published count.

• Audio length — 1 hour in one pass vs file-based with chunking handled by the API.

• Post-processing — prompt-controlled verbatim or written form vs free-form context and keyword hints, no cleanup pass.

What GPT Transcribe does better than the open checkpoint

Three things, and they are not trivial. The first is that it is an API. There is no 30B-parameter checkpoint to size, no NVFP4 build to wait for, no vLLM recipe to pin. OpenAI reports roughly 34× real-time processing, which is slower than NVIDIA's claimed throughput on a Blackwell card and irrelevant to most teams, because 34× real time is already faster than any queue they run.

The second is context handling. GPT Transcribe accepts free-form context, keyword and vocabulary hints, and multiple language hints, and returns a detected-languages array alongside the transcript. That combination is genuinely useful on code-switched audio and on domain jargon, and it is a prompting interface rather than a fine-tune — which means a product team can tune it without owning a GPU. NVIDIA's equivalent, an expected-entity list supplied in the prompt, is credited with a six-fold reduction in named-entity errors on ContextASR-Bench, but that figure is vendor-reported and unreplicated.

The third is integration surface: the same endpoint serves file transcription, streamed file transcripts, committed realtime turns over WebSocket, the Responses API and Batch. A team already on OpenAI's platform gets transcription as a parameter, not a deployment.

OpenAI's own published quality claims are broad and hard to check: a 19.27% word error rate on a 22-language Common Voice set against Whisper's 40.37%, and 8.98% on a nine-language real-world benchmark against 15.21%. Those are vendor numbers against a 2024 baseline, not against a 2026 competitor. The independent picture is friendlier to GPT Transcribe than the AA board suggests: on Perit's human-reviewed English set it lands at 1.76% WER, rank six of 22, with 61.3% of clips transcribed perfectly. That is a different test set from Artificial Analysis's, and the numbers are not comparable across the two boards — but the direction is consistent, and it puts OpenAI's model ahead of the hosted Parakeet endpoint on the same set.

Where the missing features actually bite

If your output is plain text for a search index, GPT Transcribe's gaps cost you nothing. The moment you need more, the work multiplies. Word timing means a second pass through whisper-1 or a forced aligner, which reintroduces the exact failure mode that alignment exists to fix: a boundary-shifted transcript that reads fine and breaks subtitles. Speaker labels mean a third model, joined at segment boundaries, with its own error rate to debug. And neither of those second passes knows about the other, so a two-speaker call becomes three models and a merge step.

The cascade is not a criticism of OpenAI's engineering. It is what the product is: a high-accuracy transcriber with a deliberately narrow output. Microsoft and Google both ship the same split — diarization as a separate SKU or a separate mode with its own limits. NVIDIA's bet is the opposite, and the honest way to state it is that Nemotron 3.5 Transcribe costs more compute to avoid a pipeline.

What that compute buys is documented but vendor-reported: 18.26% diarization error rate across 21 languages on the David AI benchmark, best of 14 systems; 32.4% concatenated min-permutation WER on VoiceArena, also best of 14; 8.32% of reference words dropped on overlapping English speech against a median near 31%. There is no OpenAI diarization number to set against these, because OpenAI does not publish one for this model.

Two ways to stop paying per minute

GPT Transcribe bills $0.0045 per minute, which is $4.50 per thousand minutes and about $270 per thousand audio hours. Nemotron 3.5 Transcribe bills nothing per minute and instead asks for a GPU. NVIDIA's demonstration configuration is one RTX PRO 6000 Blackwell running the NVFP4 checkpoint with a DFlash2 draft model, and the throughput headline — 100 hours of audio in under three minutes — was measured on 30-second clips at high concurrency, not on hour-long meetings. Read it as a ceiling on well-batched work.

That leaves a crossover determined by utilisation rather than by list price. A card you already own and keep busy beats $270 per thousand hours comfortably. A card you would have to buy for a queue of forty hours a month does not, and no amount of parameter efficiency changes that arithmetic.

There is a third option that the two-way framing hides, and it is the one most teams actually take: keep the managed API for the bulk, and route the hard files — the six-person meetings, the accented clinical interviews, the audio you are not allowed to upload — to a self-hosted checkpoint. That works only if the two outputs land in the same shape, which is the argument for standardising on a transcript schema with speaker and timing fields even when your current vendor does not populate them. Standardise the schema first; the model becomes swappable later.

On the layer above the transcript, OrcaRouter's job is to keep that swappability from stopping at the model boundary. The summariser, the classifier and the compliance checker that read these transcripts can all sit behind one key across 200+ models, with provider list prices passed through at 0% markup, automatic failover when an upstream stumbles, and a routing DSL that sends each workload to a different model without a second integration. GPT Transcribe and Nemotron 3.5 Transcribe are both outside that catalogue — we do not serve either — so the speech call stays wherever you buy it, and only the downstream half becomes portable.

The short version, by situation

If you want a transcript and nothing else, and you want it next week without provisioning anything, GPT Transcribe is the lower-risk choice: rank six on the independent Perit set, an integration surface you probably already have, and a price that does not require a capital purchase. If your transcripts feed subtitles, meeting analytics, speaker-attributed search or a compliance archive, the extra models you would bolt onto GPT Transcribe cost more in engineering time than the checkpoint costs in GPU hours, and Nemotron 3.5 Transcribe does that work in one pass.

The caveat that belongs on both: this is a model that is one day old, its headline benchmarks are its vendor's own, and the independent boards have not yet scored the 30-billion-parameter checkpoint at all. Do not confuse it with nvidia/nemotron-3.5-asr-streaming-0.6b, the 600M streaming model from June that does appear on independent boards and does considerably worse (3.35% WER, rank 21 of 22 on Perit). They share a family name and nothing else.