NV-Reason-CT vs GPT-5.6 Sol: 13,824 Visual Tokens Before You Type a Word

Every single chest CT you send to NV-Reason-CT costs 13,824 visual tokens before the prompt starts. That is not a quirk or a tuning choice — it is the entire design. The model's native 3D vision encoder takes a 384×384×384-mm volume and turns it into a 24×24×24 grid, and the model card is explicit that all 13,824 tokens and their three-dimensional coordinates reach the language model "without spatial downsampling." Compare that to what you are probably already paying for. GPT-5.6 Sol accepts text, images and files, and an image sent to it is a 2D picture — a slice, a screenshot of a DICOM viewer, a radiology figure pasted out of a PDF. One of these models has a volumetric pathway. The other has a context window. Most comparisons between them collapse on that distinction, and the collapse is the interesting part.

The release nobody announced

NVIDIA has not published a word about NV-Reason-CT on its newsroom. There is no launch post, no keynote slide, no press release. What exists is a Hugging Face repository created on September 8, 2026, a GitHub repository under the NVIDIA-Medtech organisation created September 2, a Gradio demo Space, and an arXiv preprint — NV-Reason-CT: 3D Visual Language Model for CT Analysis — submitted September 23, 2026 by a sixteen-author team from NVIDIA with co-authors at the NIH and the University of Zurich.

That is a real pattern, and it is worth naming precisely rather than treating as a mystery. NVIDIA's medical-imaging group releases weights quietly and lets the paper and the repository do the announcing. The predecessor, NV-Reason-CXR-3B, went out in October 2025 the same way. The difference here is scale: this is the first time that group has extended the recipe from 2D X-rays to native 3D volumes, and the paper is two days old.

What is knowable from the repository: architecture, licence, training data, the benchmark table. What is not yet confirmed: whether NVIDIA intends this as a supported product line, whether more checkpoints follow, and how it holds up under anyone's evaluation other than the authors'. The repository is at version 0.1 and describes itself as research-and-education only.

What each model actually accepts

This is where a head-to-head gets honest fast. The two models do not share an input surface.

NV-Reason-CT reads NIfTI volumes — .nii or .nii.gz — with voxel intensities in Hounsfield units. It crops automatically to the chest or the abdomen using lung Hounsfield units and a 3D morphology heuristic, resamples to 2-mm isotropic resolution, and only accepts "chest" or "abdomen" as anatomy values. It outputs text: a structured report, an answer, or a step-by-step reasoning trace, with a switch to turn the reasoning off for short answers.

GPT-5.6 Sol reads text, images and files, with a 1,050,000-token context window and up to 128,000 output tokens in a single response. It has no volumetric encoder. Feed it a CT and you must first decide how to flatten three hundred-odd slices into something a 2D image encoder will accept — a montage, a handful of key slices, a rendered maximum-intensity projection. Every one of those choices throws away the spatial relationships the 3D pathway was built to preserve.

So the honest framing is not "which model is smarter." It is that Sol's image pathway and NV-Reason-CT's 3D pathway are answering different questions. Sol can reason about a radiologist's description of a scan. NV-Reason-CT can reason about the scan.

One benchmark exists, and only one of them has a number on it

NV-Reason-CT reports Macro-F1 0.614 and Macro-AUROC 0.871 on CT-RATE's 18-label classification task, using a direct Yes/No prompt with no classification head and no task-specific adaptation. Those are the authors' own numbers from the model card, and the comparison row is the useful part: VoxelFM at 0.581 Macro-F1, Pillar-0 at 0.544, ClinFusion-8B at 0.442, CT-CLIP at 0.398, Merlin at 0.358, MedGemma 1.5 at 0.303. On the same fixed uniform threshold, a generative 3D model beats the 3D contrastive pretraining baselines and the slice-based frontier model.

One number in that table deserves scepticism rather than repetition: the AUROC gap. NV-Reason-CT reports 0.871 against VoxelFM's 0.870. One thousandth of a point, on a vendor-run evaluation, is not a win — it is a tie within whatever noise the evaluation carries. The Macro-F1 gap of 0.033 is the substantiated claim.

GPT-5.6 Sol has no CT-RATE number because CT-RATE is not a benchmark it can be run on. It has an Artificial Analysis Intelligence Index of 47, an AA-measured GPQA Diamond score of 94.1, and a Humanity's Last Exam score of 49.5 — all general-reasoning instruments. Putting 0.614 Macro-F1 beside 47 on the AA Index would be arithmetic on two different rulers. I am not going to do it, and you should distrust anyone who does.

Reasoning traces, two different products

Both models emit reasoning. That is the one genuine philosophical overlap, and the difference is instructive.

GPT-5.6 Sol exposes configurable reasoning effort, so you trade latency and cost against depth per request, and you can request the reasoning back through include_reasoning. It is a general capability applied to whatever you send it.

NV-Reason-CT's reasoning is narrow on purpose. The training corpus is roughly 550,000 structured QA examples drawn from 70,111 unique CT volumes, and a portion of it is radiologist-authored reasoning collected from recorded and transcribed expert interpretations. The GRPO stage that follows SFT uses verifiable rewards over chest and abdominal abnormality sets — meaning the reward signal is grounded in whether a stated finding is actually present in the volume, not in whether the prose reads well. The model card describes the output as "reviewable observations, differential diagnoses, and uncertainty," and it carries an explicit caution that generated reasoning "is not guaranteed to represent the model's internal computation." That caution is doing real work. It is the difference between a trace you can check against the data and a trace you can only admire.

The size and the licence, in one pass

• Parameters — NV-Reason-CT 4.69B total / 4.35B active in the CT pathway, from a Qwen3.5-4B backbone plus a Primus 3D ViT vs GPT-5.6 Sol undisclosed • Checkpoint size — NV-Reason-CT ~10.6 GB BF16 safetensors, runs on Ampere/Hopper/Lovelace vs GPT-5.6 Sol API-only • Input — 3D NIfTI CT volumes in Hounsfield units vs text, image, file • Context — NV-Reason-CT not applicable, one volume is one fixed 13,824-token prefix vs Sol 1,050,000 tokens in, 128,000 out • Licence — OpenMDW-1.1, weights downloadable vs proprietary, API access only • Availability — vendor repository and its own weights vs routed on OrcaRouter at $4.00 / $20.00 per million, with Sol's rate above 272K input tokens doubling to $8.00 / $30.00

What the split actually costs you

The cost comparison only makes sense once you accept that the workloads are different, so here is the version that does make sense.

Sol is billed per token and its price has a cliff: $4.00 per million input and $20.00 per million output up to 272K tokens of prompt, then $8.00 and $30.00. On OrcaRouter it is served at OpenAI's own list price with zero markup, which matters more than it sounds — the rate you see is the rate OpenAI publishes, so any price move reaches your bill the same day rather than at the next contract renewal. Its cache-read rate is $0.40 per million, a tenth of input, which is what makes long agentic loops with a large fixed prefix arithmetically survivable.

NV-Reason-CT has no per-token price at all. You download 10.6 GB and pay for the GPU. That is a capex-versus-opex question, and it only has one answer: at high enough volume, self-hosting a 4.35B active-parameter model wins on cost. Below that volume it does not, and the crossover point depends entirely on your GPU utilisation. Nobody outside NVIDIA has published a throughput figure for this checkpoint yet, so anyone quoting you a crossover point is guessing.

What a router does help with here is the part of the workflow that is not the scan. A production clinical-research pipeline is rarely one model — it is extraction, a reasoning step, a summarisation step, sometimes a second opinion. OrcaRouter exposes 200+ models behind one OpenAI-compatible endpoint with automatic failover across providers, so the general-purpose half of that pipeline can be swapped or rerouted without a second contract. NV-Reason-CT is not one of the models we route; it is a checkpoint you run yourself. The two halves of the pipeline can still sit behind one key.

The open question

NV-Reason-CT is two days old as a paper and seventeen days old as a repository, and the field it belongs to is small enough that the next data point will be informative. Watch for three things: an independent CT-RATE reproduction, a second checkpoint in the family, and any sign that NVIDIA's medical group is treating this as a line rather than a paper. Until then, the defensible position is narrow and clear — this is the strongest native-3D CT reasoning checkpoint currently downloadable, judged on the authors' own table, and it is not a replacement for a general frontier model at anything except reading a CT.