NV-Reason-CT vs Grok 4.6: Who Reads the Scan, and Who Reads the Report

There are two ways to put an AI on a CT pipeline and only one of them is about the image. NV-Reason-CT is the first: a 4.69B-parameter 3D vision-language model that reads NIfTI volumes directly and was published to Hugging Face on September 8, 2026 without a launch post from NVIDIA. Grok 4.6 is the second: a 500,000-token, text-and-image model that shipped on August 12, 2026, costs $2.00 per million input tokens, and is very good at the part of the job that happens after somebody has already written the finding down. Put them in a comparison and the usual question — which one is better — turns out to be malformed. They are not competing for the same slot in the pipeline. They are competing for the same budget line.

What actually shipped on each side

NV-Reason-CT arrived as a repository, a paper and a demo Space, not as a product. The GitHub repository under NVIDIA-Medtech was created September 2, 2026; the Hugging Face weights followed on September 8; the arXiv preprint, NV-Reason-CT: 3D Visual Language Model for CT Analysis, landed September 23 with sixteen authors from NVIDIA, the NIH and the University of Zurich. NVIDIA's newsroom carries nothing about it. The model card describes version 0.1 and restricts use to research and education.

Grok 4.6 is the opposite kind of release. It is a first-class commercial endpoint, listed as "Latest" in the vendor's developer documentation, priced on a published rate card, and available as an OpenAI-compatible model that existing integrations adopt by changing a base URL. It succeeded Grok 4.5 with the same context window, the same input surface, and the same base pricing.

That asymmetry is the whole story of this comparison. One is a research artifact that happens to be downloadable. The other is infrastructure. Reading them against each other tells you where the line between "research artifact" and "infrastructure" currently falls in medical imaging — and it is not where you would guess.

The division of labour, in inputs and outputs

• Volumetric input — NV-Reason-CT reads .nii/.nii.gz NIfTI volumes in Hounsfield units, chest or abdomen only vs Grok 4.6 reads text, images and files, no volumetric pathway • Spatial handling — NV-Reason-CT converts a 384×384×384-mm volume to a 24×24×24 grid of 13,824 tokens with 3D MRoPE, no downsampling vs Grok 4.6 treats an image as a 2D picture • Context — NV-Reason-CT one volume is one fixed prefix, no context window in the usual sense vs Grok 4.6 500,000 tokens • Output — NV-Reason-CT structured report, answer, or reasoning trace with thinking disableable vs Grok 4.6 text with structured outputs and tool calls • Licence — NV-Reason-CT OpenMDW-1.1, 10.6 GB BF16 weights vs Grok 4.6 proprietary, API only • Price — NV-Reason-CT GPU capex, no per-token rate vs Grok 4.6 $2.00 / $6.00 per million, doubled above 200K input

The pair reads as a natural handoff. NV-Reason-CT produces the finding from the volume; Grok 4.6 is the model you would hand a thousand of those findings to, in one context window, to look for patterns across a cohort. Nobody has published that pipeline. It is the obvious architecture, and it is worth saying plainly that it is untested.

Grok 4.6's numbers, and whose they are

Two figures for Grok 4.6 circulate together and they are not the same kind of thing. The GPQA Diamond score of 94.9 is measured by Artificial Analysis, not reported by the vendor — which is the more trustworthy of the two categories precisely because a third party ran it. The Artificial Analysis Intelligence Index score of 44, on index revision v4.3.2, places Grok 4.6 at rank 28 of 211 in its class. Artificial Analysis also records the model as proprietary: the weights are not publicly available.

On the same board, AA measures Terminal-Bench 2.1 at 88.4, SciCode at 56.5, Humanity's Last Exam at 42.9, 𝜏²-banking at 50.7, and long-context recall at 80.3. Those are general-reasoning instruments. Not one of them can be run on a 3D CT volume, and that is not a knock on Grok 4.6 — it is a statement about what the benchmark suite measures.

The CT side has exactly one table, and it is the authors'

NV-Reason-CT's entire public evidence base is one classification table on CT-RATE's 18 labels, at a fixed uniform threshold, using a direct Yes/No prompt with no classification head. It reports Macro-F1 0.614 and Macro-AUROC 0.871, against VoxelFM at 0.581/0.870, Pillar-0 at 0.544/0.861, ClinFusion-8B at 0.442, CT-CLIP at 0.398/0.733, Merlin at 0.358/0.662, and MedGemma 1.5 at 0.303.

These are vendor-reported, from the model card, and I am labelling them as such because no independent party has reproduced them yet. Two things are worth noticing inside the table. The F1 margin over VoxelFM is 0.033, which is a real gap on 18 labels with a fixed threshold. The AUROC margin over the same model is 0.001, which is a tie. Both numbers appear in the same row of the same table, and only one of them carries a claim.

The other thing the table tells you is about the paper's own framing. The comparison models are 3D contrastive pretraining systems, a fused 2D/3D generative MLLM, and a slice-based frontier model. The authors chose to benchmark against systems that ingest volumes, because the only meaningful claim here is that a generative 3D model can beat discriminative 3D models at classification without a head. It says nothing about how NV-Reason-CT compares to a general frontier model at anything, because that comparison does not exist on this axis.

Where the money goes, and why the tier boundary matters

Grok 4.6's rate card has a detail that quietly decides a lot of architecture: prompts above 200K input tokens bill at double the base rate — $4.00 in, $12.00 out, with the cache-read rate rising from $0.50 to $1.00. A cohort-review job that accumulates findings in one long context is exactly the workload that crosses that line. Under it, the cache-read rate of $0.50 per million is a quarter of the input price, which is what makes looping a large fixed prefix over thousands of reports affordable rather than merely possible.

Through OrcaRouter, Grok 4.6 is served at that provider list price with zero markup, so the tier arithmetic above is the arithmetic you actually pay, and any future adjustment to it reaches your bill the same day rather than at the next renewal. The reason to route a job like this rather than hard-code one endpoint is that the cohort-review half of a clinical pipeline is where you will want to try three different models before settling — and on one OpenAI-compatible key with automatic failover across providers, that experiment costs a model-name change.

The CT half of the pipeline has no such option. NV-Reason-CT is not hosted on OrcaRouter — it is a 10.6 GB checkpoint with custom model code that you run yourself, on Ampere, Hopper or Lovelace silicon, with trust_remote_code=True. We are not going to imply otherwise. If you are building this pipeline you are buying GPUs for one half and tokens for the other, and the fact that both halves can at least be managed from one console is the only integration claim worth making.

What a second reader is actually worth

The NV-Reason-CT paper includes a preliminary expert-radiologist study reporting "favorable confidence ratings for AI-assisted review" and a 50% reduction in average reported interpretation and reporting time. Those are strong numbers and they come with a caveat the paper states itself: preliminary, and the authors' own study design. There is no published independent replication, and a 50% time reduction in a controlled reading study is exactly the kind of result that shrinks under replication. It is worth tracking. It is not worth quoting as established.

Read against that, Grok 4.6's contribution to a radiology workflow is not diagnosis at all. It is the layer that reads the reports — the triage queue, the follow-up letter, the coding, the prior-authorisation packet. That work is text, it is high volume, it is cheap per unit, and the failure mode of getting it wrong is administrative rather than clinical. It is also where a 500K context window and a $0.50 cache read genuinely earn their place.

So the split this comparison lands on is blunt: NV-Reason-CT has a real 3D pathway and no distribution, no price, and no independent verification. Grok 4.6 has distribution, a price, independent benchmark coverage, and no volumetric pathway at all. If you need the scan read, only one of these can do it. If you need the paperwork around the scan handled, only the other one is a product.