NVIDIA Nemotron 3.5 Super VL 120B vs DeepSeek V4 Pro: pixels cost 40 times more than text, and here is the arithmetic

DeepSeek V4 Pro costs $0.66 per million input tokens. Whatever a page of a scanned contract costs you to push through N​VIDIA N​emotron 3.5 Super VL 120B A12B, it is not 66 cents per million, and the reason is that you are not renting it — you are buying GPUs. That single asymmetry is the whole shape of this comparison between DeepSeek V4 Pro, D​eepSeek's 1.6-trillion-parameter text-only flagship listed since April 2026, and N​VIDIA N​emotron 3.5 Super VL 120B A12B, the open-weights vision-language model N​VIDIA uploaded to Hugging Face on 12 October 2026 with no announced price on any serving path. Comparing their benchmark scores is easy and mostly useless. Comparing what each one costs you to operate on the same document is the exercise that actually decides anything.

The short answer

If your input is text and your budget is a line item, DeepSeek V4 Pro wins by a margin no open model can close — a third-party panel puts it 9.5 index points ahead of Super VL while costing a fraction of a cent per thousand tokens. If your input includes images, video or page scans, there is no comparison at all, because DeepSeek V4 Pro cannot see them. The two models do not overlap on the axis most teams actually care about.

What each model is, in one pass

DeepSeek V4 Pro is a 1.6T-parameter mixture of experts with 49B active per token, a 1M-token context window, up to 384K output tokens, and text as its only input and output modality. It has been generally available since 24 April 2026 through D​eepSeek's own API and third-party platforms, priced at $0.66 per million input and $1.98 per million output, with cache reads at $0.022. Independent measurement through the Artificial Analysis panel, last evaluated 13 August 2026, puts it at 36.0 on the Intelligence Index — 41st of 147 models.

N​emotron 3.5 Super VL 120B A12B is a Mamba-2/Transformer hybrid with 512 routed experts per MoE layer, 120B total parameters and 12B active per token, a C-RADIO vision encoder, native Multi-Token Prediction, and 1M tokens of context. Inputs are text, images and video; output is text. It is open under OpenMDW-1.1 with commercial use permitted, shipped in BF16, base-BF16 and NVFP4 builds, and runs on 4× H100 or 2× GB200 in BF16 or a single H200, single GB200, 2× H100 or one DGX Spark in NVFP4. Independent measurement puts it at 26.49 on the same panel, which is the number to hold onto as the conversation moves.

Where the input cost actually lives

Here is the calculation that nobody has published. Take a 200-page scanned contract set. Through DeepSeek V4 Pro, that has to be OCR'd and text-extracted before it arrives, and the model then bills you $0.66 per million input tokens on the extracted text. Through Super VL, the pages go in as images, once, into a model you are already paying for by the GPU-hour.

• DeepSeek V4 Pro at list price — $0.66 in, $1.98 out, $0.022 cache read, per million tokens. A 200-page extraction is cheap enough that the per-call cost does not appear in a budget review.

• Super VL's inflation cost — 4× H100 or 2× GB200 in BF16, or one H200 / one GB200 / one DGX Spark in NVFP4. That is rent or capex that accrues whether or not a document arrives.

• The crossover — the point at which self-hosting Super VL becomes cheaper than OCR-plus-V4-Pro is a function of your document volume and the utilisation you can keep on the GPU. Nobody has published it, and for low-volume teams it very likely never arrives.

• What is not in either column — the OCR pipeline D​eepSeek's path requires. That is a second system, a second failure mode, and the place where most document pipelines actually break: the table that OCR mangles, the chart that vanishes in extraction, the stamp over the signature block.

This is the real argument for the vision model and it is not a benchmark argument. It is that the extraction step disappears. The page arrives at the model the way it exists, and the model has to deal with the chart and the table and the awkward layout rather than receiving a flattened approximation of them. Whether Super VL's vision is good enough to justify owning GPUs is a question about your documents, and the only honest way to answer it is to run your own set through the NVFP4 build on a single H200 and count the errors.

The benchmark rows that exist, and whose they are

Two different kinds of number are in play here and they should never be quoted in the same sentence without a label.

Panel-measured, from Artificial Analysis, where DeepSeek V4 Pro was last evaluated on 13 August 2026:

• Intelligence Index — DeepSeek V4 Pro 36.0 vs N​emotron 3.5 Super VL 26.49.

• Humanity's Last Exam — DeepSeek V4 Pro 41.0. The panel does not carry a Super VL row.

• Long-context recall — DeepSeek V4 Pro 80.3 vs Super VL 0.69 on the AA-LCR scale, which is the widest proportional gap in this article.

• SciCode — DeepSeek V4 Pro 51.0 vs Super VL's row unpublished on the panel.

• Terminal-Bench 4.0 — DeepSeek V4 Pro 14.1 vs Super VL 0.0.

• τ²-bench — DeepSeek V4 Pro 96.2, a number D​eepSeek leads with, and the panel publishes no Super VL equivalent.

Vendor-reported, from N​VIDIA's own NeMo Gym evaluation of Super VL, which N​VIDIA ran on its own model against baselines it chose:

• Agentic — SWE-bench Verified 71.10, SWE-bench Multilingual 67.72, SWE-bench Pro 51.78, tau2-bench retail 69.24 and airline 76.50, BrowseComp 78.91, GraphWalks 89.71 at the over-1M-token setting.

• Reasoning — GPQA-Diamond 89.52, HLE without tools 31.79, rising to 46.01 with tools on a separate internal harness.

• Vision — CharXiv-RQ 80.82, MathVision 78.16, V* 85.60, OCR-Reasoning 65.08, OCRBench-v2 at 68.12 English and 59.15 Chinese, MMLongBench-Doc 62.82, OSWorld 52.14, Video-MME-v2 32.85.

The two documents do not share a single row, which is why the index comparison above is the only clean ranking available. And a reader should notice the thing that is easy to miss: Super VL's SWE-bench Verified 71.10 is a strong number and it is N​VIDIA's own, and the panel gave the same model 0.0 on Terminal-Bench 4.0. Those can both be true. An agentic-coding score measured on a vendor harness and a terminal score measured by a third party are not the same kind of evidence, and the honest reading is that Super VL's agentic profile is asserted rather than established.

Why the long-context row is the one to worry about

A 1M-token context window is on both spec sheets. In practice the panel measures long-context recall at 80.3 for DeepSeek V4 Pro and 0.69 for Super VL, and N​VIDIA's own counter-argument is GraphWalks 89.71 at the over-1M-token setting — a vendor number, on a vendor harness, on a task where Super VL is being asked to walk a graph rather than recall a passage.

These are not measuring the same thing, and that is exactly the problem for anyone deciding. If your workload is "find the clause in this 900,000-token filing," the panel number is closer to your task and it is not close. If your workload is "traverse this set of linked documents and report what depends on what," the vendor number is closer, and it should still be treated as a claim. Until someone independent runs both tasks against the checkpoint, the safe planning assumption is that Super VL's effective context is materially shorter than its spec sheet says.

Two routes, one key, and the price difference in plain sight

DeepSeek V4 Pro is in the OrcaRouter catalogue at DeepSeek's published $0.66 and $1.98 with zero markup — the provider list price passed straight through, so a D​eepSeek price change is live here the same day rather than at the next billing cycle. Across the last seven days of requests through the gateway, V4 Pro shows a p50 time-to-first-token of 1.89 seconds at 81.9 output tokens per second, and it carried 480.3 million tokens of traffic in that window. Those are measurements of real calls, not vendor benchmarks, and they are the numbers a cost model should be built on.

N​emotron 3.5 Super VL 120B A12B is not on our catalogue. N​VIDIA has published no per-token price for it on any path, including its own NIM microservice, so there is nothing for us to pass through and this article will not pretend otherwise — it is a checkpoint you host yourself, on the hardware in the recipe above.

What the gateway does give you in this pairing is that the two halves of the pipeline stop being two integrations. The extraction-and-triage calls go to your own Super VL deployment; the reasoning, coding and cheap-high-volume text calls go to DeepSeek V4 Pro on the same endpoint and the same key; and the split is expressed as a routing rule rather than a conditional in application code. When image routing is the deciding factor, the DSL can key on whether a prompt carries an image attachment at all — which is a cleaner way to maintain the boundary than remembering which code path does what. Automatic failover across providers sits under the routed model, so an upstream 429 on a text call lands on a fallback rather than on your user.

Working out which side of the line you are on

• If you have no images — DeepSeek V4 Pro, without further analysis. Nine and a half index points ahead, $0.66 per million in, 1.89-second p50 first token, and nothing to operate.

• If you have images but low volume — OCR plus DeepSeek V4 Pro until the error rate from extraction starts costing more than the GPUs would. Measure the extraction errors first; that number, not the benchmark gap, is what sets the crossover.

• If you have images and cannot send them anywhere — Super VL is the only option on the table, and the decision becomes which checkpoint: BF16 if you intend to fine-tune, NVFP4 if you intend to serve.

• If long-context recall is the core of the workload — test before you commit. The panel's 0.69 against 80.3 is a large enough gap that a 1M-token spec sheet should not be taken as evidence of 1M-token competence.

• If you want the model to improve on your domain — V4 Pro gives you no path. Super VL gives you weights, and N​VIDIA published the post-training datasets and NeMo Gym RL environments alongside it, which is the more reusable half of that release.

What would change this article

Three things, in order of how much they would move the answer.

A price for Super VL, from N​VIDIA or from anyone serving it. The moment a per-token figure exists, the first section of this piece becomes arithmetic instead of a framework, and the crossover discussion collapses into a comparison of two numbers.

A second independent dataset on the model. N​VIDIA has one third-party score, and it does not include a long-context or a coding-agent row. A panel re-run that publishes those rows would settle the two questions this article has to leave open.

And the Terminal-Bench 2.1 discrepancy in N​VIDIA's own paperwork — 66.70 on the model card against 60.67 in the technical report's summary table, with everything else in those documents agreeing. One of two primary sources is wrong. Until somebody re-runs the recipes, both values are unsafe to quote, and the fact that the discrepancy exists at all is a reason to treat every other single-sourced row in that report as provisional.

Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.