NV-Reason-CT vs DeepSeek V4 Pro: The Cost of Reading a Scan, and When to Run the Batch

Here is an operational fact that shapes more clinical pipeline budgets than any benchmark: DeepSeek V4 Pro costs $0.66 per million input tokens and $1.98 per million output tokens, and those rates double during two windows each day, 01:00 to 04:00 and 06:00 to 10:00 UTC. Batch work that runs outside those windows costs half what it costs inside them. Meanwhile NV-Reason-CT, NVIDIA's 4.69-billion-parameter 3D CT reasoner, has no rate card at all — it is a set of weights you download and run, with no published throughput figure for a single 13,824-token study. One of these models you schedule. The other you have to measure before you can budget it.

That asymmetry is the useful way into this comparison, because the two models occupy opposite ends of the pipeline and fail in opposite ways financially. NV-Reason-CT is a fixed-cost instrument: the marginal study is nearly free once the hardware is bought, but the hardware decision is large and the per-study latency is undocumented. DeepSeek V4 Pro is a variable-cost engine: nothing to buy, everything metered, and the meter has a schedule you can read off the calendar.

What each one is, in one line each

• Job — NV-Reason-CT reads a chest or abdominal CT volume and emits structured findings; DeepSeek V4 Pro reads and writes text

• Architecture — a 3D vision transformer called Primus initialised from COLIPRI, with a Qwen3.5-4B language backbone and 3D multi-axis rotary position embedding, at 4.69B parameters of which 4.35B are active; against a 1.6-trillion-parameter mixture of experts with 49 billion active per token

• Input — one-channel NIfTI volumes (.nii, .nii.gz) in Hounsfield units, LPS-oriented, resampled to 2 mm isotropic and cropped to anatomy; against text

• Context — a single volume becomes 13,824 visual tokens in a 24 x 24 x 24 grid, with no spatial downsampling and no merging layer; against 1,048,576 tokens in and 384,000 out

• Anatomies supported — chest and abdomen, and nothing else; against everything text can describe

• Price — no list price, self-hosted, no published per-study throughput; against $0.66 / $1.98 per million tokens, doubling in the two UTC windows named above, with cache reads at $0.022

• Measured latency — not published; against a 3,483-millisecond median time to first token at 96.01 output tokens per second, with a 0.75% error rate

Two rows there are worth more than a line each. The context row is the obvious one — a million tokens against 13,824 — but the units are not comparable and it is a mistake to read them as a capability gap in either direction. 13,824 tokens is what one CT study costs to represent faithfully. A million tokens is how much surrounding text you can hold. The first number is a statement about resolution; the second is a statement about memory.

The latency row is the one that gets skipped. DeepSeek V4 Pro's 3.48-second median to first token and 96 tokens per second are measured, published figures, and they let you size a text job on paper. NV-Reason-CT's absence of any equivalent figure is not a criticism of the model; it is the reason a CT pipeline cannot be costed before someone runs it. A 4.69B model over 13,824 visual tokens is not a small computation, and until you have timed it on your own hardware you are guessing.

Reading the two benchmark records without fooling yourself

DeepSeek V4 Pro's record is a mix of independent measurement and vendor reporting, and the mixture is instructive because it contains one number that looks alarming and is not.

• Artificial Analysis Intelligence Index — 36, which sits at the 72.4th percentile of models measured

• GPQA Diamond — 92.8, which is near the top of the field

• Humanity's Last Exam — 41, and 41 on MMLU-Pro, 62.51 on the maths index

• τ²-Bench — 96.20, with IFBench at 76.46 and tau_banking at 39.59

• Long-context recall — 80.33

• SciCode and Terminal-Bench — 51 and 78.65 on the second version of Terminal-Bench

A composite index of 36 sitting beside a GPQA Diamond of 92.8 reads like a contradiction until you notice what an index is. It is an aggregate across many evaluations, and a model that is excellent at a handful of them and merely adequate at others will place mid-field on the sum while topping individual boards. Anyone quoting the 36 alone to argue DeepSeek V4 Pro is mid-tier is quoting an average as if it were a maximum. Anyone quoting the 92.8 alone to argue it leads the field is quoting a maximum as if it were an average. Both numbers are from Artificial Analysis, and both are true.

NV-Reason-CT's record has no such mixture, and that is the honest problem with it. Its CT-RATE figures — 0.614 F1 and 0.871 AUROC across eighteen labels, using a fixed uniform threshold and a direct yes/no prompt with no classification head and no task-specific adaptation — come from NVIDIA's own paper and nowhere else. The comparison set in that paper (VoxelFM at 0.581, 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) is entirely other imaging models. Not one general text model appears in it, and no third party has reproduced any of the numbers.

The scheduling question, worked through

The timed pricing on DeepSeek V4 Pro is the most operationally actionable thing about either model, so it deserves a concrete walk-through.

A realistic text-side workload for a CT programme is not glamorous. It is extracting structured fields from a corpus of historical reports so you have labels to train against, converting DICOM directory trees to NIfTI at scale, writing and re-writing the evaluation harness, normalising units and terminology across four datasets, and summarising cohorts. Call it a hundred million input tokens and ten million output tokens for a mid-sized programme, which is a deliberately round figure standing in for "a lot of text work".

• Inside a doubled window — $1.32 and $3.96 per million, so $132 plus $39.60 on those volumes

• Outside those windows — $0.66 and $1.98 per million, so $66 plus $19.80

• The difference — roughly $86, or 49% of the off-peak bill, for moving a job that is batch-able by nature

• Cached reads — $0.022 per million, which is a sixtieth of the input rate, so any prompt prefix you reuse across the corpus is close to free

That is not a rounding error, and it is available because the work is asynchronous. Report extraction does not need to happen at 14:00. Nothing about a DICOM conversion job cares what time it is. The pricing structure is a direct invitation to schedule, and a pipeline that ignores it is paying a 49% premium for the privilege of running when it felt like it. Cache reads matter for the same reason: a shared system prompt or a fixed extraction schema, reused across thousands of reports, sits at a sixtieth of the input rate.

NV-Reason-CT has no equivalent lever, because it has no meter. The cost of reading a scan is whatever your GPU costs per hour divided by however many scans per hour you can push through it, and the second number is the one nobody has published. If a single study takes two seconds, a single L40S handles a large programme comfortably. If it takes twenty, the hardware arithmetic changes completely. NVIDIA tested on H100 and L40S, which tells you the class of card, not the rate.

The trap in assuming they can be substituted

The mistake this matchup invites is subtler than the usual category error, because there is a version of it that looks reasonable on a slide.

DeepSeek V4 Pro holds 1,048,576 tokens and emits up to 384,000. A CT volume is, by the reckoning of the model that reads it properly, only 13,824 tokens. So a text model with a million-token window has room for seventy-odd CT studies at that token count. The arithmetic is correct and the conclusion — that you could feed CT data to a text model and save yourself the imaging infrastructure — is wrong, for the reason the 13,824 figure exists in the first place.

That number is not a compressed summary of a scan. It is the scan, at 2 mm isotropic resolution over a 384-millimetre cube, cut into 8 x 8 x 8 patches, handed to the language model with no spatial downsampling and no merging layer. The token count is high precisely because nothing was thrown away: the inter-slice spacing that makes a nodule measurable, the density values that make a texture a threshold rather than an adjective. A text model cannot ingest those tokens as volumes any more than a document can. It has the budget. It does not have the interface.

The paper's own internal comparison puts a number on the difference. MedGemma 1.5, given up to 85 axial slices — the best a two-dimensional or slice-stacked front end can reasonably do — scores 0.303 F1 against 0.614 for the native volumetric model. That gap is the value of the three-dimensional encoder, and no amount of context window closes it.

The reverse substitution fails for plainer reasons. NV-Reason-CT supports chest and abdomen, takes a NIfTI file rather than a prompt, has no tool use, and its model card restricts it to research and education and states that it is not a medical device and that outputs can be incorrect. It cannot extract fields from a report, and it cannot write the script that builds your corpus.

Where we fit, and where we deliberately do not

DeepSeek V4 Pro is served through OrcaRouter as deepseek/deepseek-v4-pro, at the provider's list rate with zero markup, inside a single API covering more than 200 models. The pass-through matters more than usual for a model with a time-of-day price: when a vendor changes a rate or a window, the rate on our side changes the same day, because there is no markup layer in between holding an old number. Automatic failover covers the case where a provider degrades mid-batch, which for a long unattended extraction job is the failure mode that actually costs you a night, and the routing DSL lets you send individual requests to the model that should handle them rather than running everything past one endpoint.

NV-Reason-CT is not on our platform. No NVIDIA model is. The CT side of this architecture is your own hardware with your own downloaded weights, and we are not going to write a sentence that lets a reader think otherwise. Given that the input is patient-derived volumetric data and the licence is OpenMDW-1.1 without any hosted service attached, local execution is where that model belongs regardless.

What the pairing actually tells you

The two models do not compete, and the point of comparing them is that they fail differently. NV-Reason-CT gives you a capability nothing else in the open does — native three-dimensional CT reasoning at the full resolution of the study — and asks you to accept an unbudgeted per-study cost, an unpublished throughput, and a licence that forbids clinical deployment. DeepSeek V4 Pro gives you a metered engine with a published latency, a published error rate, independent measurements and a price you can halve by looking at a clock, and it cannot see a scan at all.

If you are building the thing rather than buying it, the shape that survives contact is the boring one. Run the CT read locally, budget it by measuring rather than by quoting, and put everything textual around it on a rendezvous with the off-peak window. The one schedule nobody should copy is the one that runs a hundred million tokens of extraction at 02:00 UTC because that is when the batch happened to be ready.