A title card for the comparison, headlined 'Nemotron for Science vs GPT-Rosalind', with four statistics: open weights under OpenMDW-1.1 with no application, GPT-Rosalind at $5.00 input and $25.00 output per million tokens behind trusted-access review, roughly 50,000 trainable LitmusBench v0.2 chemistry questions, and 81% of GROMACS fixes at KTH in one overnight DGX Spark run.
Guides & Insights

Nemotron for Science vs GPT-Rosalind: Open Weights Against a Trusted-Access Gate

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Ask what each of these two costs and you get an answer that says almost nothing about the other. NVIDIA Nemotron for Science, announced at SC26 on 16 November 2026, is a downloadable package: the open Nemotron 3.5 Super VL checkpoint under OpenMDW-1.1, a roughly 50,000-question chemistry dataset called LitmusBench v0.2, and post-training recipes. Its marginal cost is whatever GPU you already own. GPT-Rosalind is a metered API behind an application: the company's life-sciences specialist, introduced 16 April 2026, opened worldwide to eligible organisations on 11 September 2026, billing at $5.00 per million input tokens and $25.00 per million output since 5 October. One is a kit. The other is a service with a vetting step in front of it. That difference decides more about which one your lab should use than any benchmark score on either side.

Both labels say "science," and that word is doing a lot of work in both cases. Neither model is a general-purpose scientist; each is aimed at a specific slice of research work, built by a company with a specific commercial interest in that slice, and evaluated by that company. So this comparison separates three questions that get collapsed in launch coverage: what each model can actually be pointed at, who is allowed to point it, and what one experiment costs on each.

Two different products, not two versions of one

GPT-Rosalind is a reasoning model tuned for biology — molecules, proteins, genes, pathways, disease mechanisms — with a documented plugin ecosystem of 50-plus life-sciences tools and a research-note track record that includes a distributed-agent run screening more than 700 references down to 11. It is accessed through ChatGPT, Codex and the API, and the API route requires clearing a vendor-run trusted-access programme. That is a gate on the model itself, not on a feature. The capability is real; the question is whether you get it.

Nemotron for Science is not a model in the same sense. The model inside it, Nemotron 3.5 Super VL, shipped its weights on 12 October 2026 and was announced by NVIDIA on 15 October — five weeks before SC26. What landed on 16 November is the surrounding apparatus: a chemistry evaluation set, post-training recipes, and documentation for wiring the model to CUDA-X scientific software as callable tools. Nobody gates it. You clone it, you size the hardware, you fine-tune it on your own data, and the only approval step is your own compute budget.

That asymmetry is the entire comparison, and everything below is a consequence of it.

The scoreboard, dimension by dimension

• Access model — Nemotron for Science: open weights on Hugging Face under OpenMDW-1.1, commercial and non-commercial use, no application. GPT-Rosalind: trusted-access review before the first API call; chat access through a separate academic programme that explicitly excludes API calls.

• List price — Nemotron for Science: no vendor rate card; self-hosted cost is your hardware. Third-party inference lists Super VL at $0.20 per million input and $0.50 per million output, with a 75% cache discount and a $0.13 blended rate. GPT-Rosalind: $5.00 input / $25.00 output per million, $0.50 cached input.

• Parameters — Nemotron 3.5 Super VL: 120B total, 12B active per token, Mamba-2 and Transformer hybrid MoE with 512 routed experts per MoE layer. GPT-Rosalind: undisclosed.

• Context and modality — Both 1M tokens. Super VL takes text, images and MP4 video up to two minutes and outputs text. GPT-Rosalind takes text and scientific file inputs through ChatGPT, Codex and the API; no general image-input claim is published.

• Domain evidence — Nemotron for Science: seven named partners and institutions, including 76% reaction recovery on RxnScribe at Terray, 81% of broken GROMACS setups repaired by a KTH agent in one overnight DGX Spark run, and 61% versus 54% quantum-dot synthesis accuracy against GPT-6 Astra at Lila Sciences. GPT-Rosalind: BixBench leading, MedChemBench 27.5%, GeneBench 21.6%, LabWorkBench 63.2%, and a 4.42% performance-per-token gain on LifeSciBench. All vendor-reported on both sides.

• Independent score — Nemotron 3.5 Super VL: Artificial Analysis Intelligence Index 14, 291 tokens per second, 11.85s to first answer token. GPT-Rosalind: not tracked on that index.

A two-column comparison scoreboard for Nemotron for Science and GPT-Rosalind across six dimensions: access, list price, parameters, context and modality, domain evidence, and independent score. Nemotron is open weights at 120B total and 12B active with an Artificial Analysis Intelligence Index of 14; GPT-Rosalind is trusted-access reviewed, priced at $5.00 input and $25.00 output per million with 27.5% on MedChemBench and 63.2% on LabWorkBench, and is not tracked on that index.

Chemistry and biology are not the same benchmark

This is the mismatch that a spec-by-spec reading hides. GPT-Rosalind's published evidence is overwhelmingly biological: protein-expression ranges, gene-level reasoning, medicinal chemistry, lab protocol work. Nemotron for Science's published evidence is overwhelmingly chemical and computational: reaction-scheme extraction, solubility prediction, metabolic fate, quantum-dot synthesis, molecular dynamics repair. Both are "science," and a computational chemistry group and a translational oncology group are not in the same market.

Where they genuinely overlap is the agentic layer, and here NVIDIA's evidence is unusually concrete. A model that reads a broken GROMACS input deck, works out why the topology and the coordinate file disagree, and produces a fix that the engine then accepts is doing multi-step tool use with a verifiable reward — and KTH reports 81% success on test cases they constructed. The comparable Rosalind evidence is the 32-minute cholesterol-trafficking hypothesis run from October, where thousands of cooperating agents screened 700-plus references down to 11 and blinded reviewers preferred the output in roughly three quarters of comparisons. Both are first-party. Both describe a loop that plans, executes and checks. Neither publishes a completion rate under production load, which is the number that would actually let you compare them.

Read the two evidence bases as complementary rather than competing. If your work is wet-lab biology, NVIDIA's chemistry dataset is close to irrelevant to you and Rosalind's MedChemBench and LabWorkBench rows are the ones to read. If your work is computational chemistry or molecular simulation, the GROMACS and quantum-dot results describe your actual week, and Rosalind's biology tuning is the thing you would be paying a premium for and not using.

What an experiment costs, worked out

Per-token pricing is the wrong unit for both of these, so here is the arithmetic that matters. Take a post-training run on Nemotron for Science: a 120B model at BF16 wants 4× H100 or 2× GB200, or a single DGX Spark on the NVFP4 build. A single overnight run on one DGX Spark — the KTH configuration — is a fixed, knowable cost against hardware you either own or rent by the hour. The dataset is free and the recipes are published. You can run the loop a hundred times and the marginal cost of run ninety-nine is the same as run one.

Now the same iteration on a metered specialist. At $5.00 input and $25.00 output per million tokens, a single long-context pass over a large scientific corpus is not a rounding error, and the cost of a failed attempt is charged at the same rate as a successful one. That is why cached input at $0.50 matters so much on the Rosalind side: if your prompts share a large prefix, the discount is 90% and the economics change completely. If they do not, they do not.

The honest conclusion from that arithmetic is not "open is cheaper." It is that the two pricing shapes reward different work. Iterative fine-tuning rewards open weights, because you pay for compute and nothing else. Burst evaluation of a hundred hypotheses against a frontier model rewards a metered API, because you pay for what you use and nothing else. Most research groups do both, which is the awkward part.

A two-column cost card titled 'What one experiment costs, worked out', comparing a self-hosted Nemotron run against a metered GPT-Rosalind call across unit of cost, cost of run ninety-nine, cost of a failed attempt, shared prompt prefix discount, hardware requirement, and the kind of work each pricing shape rewards.

Where the routing layer fits, and where it does not

Neither checkpoint is on our catalogue. We do not serve Nemotron 3.5 Super VL and we do not serve GPT-Rosalind, and it would be a strange article that pretended otherwise — the practical answer for the first is that you download it, and for the second that you apply to the vendor.

What does sit on our side of the line is the part of a science agent that is not the science model. The KTH GROMACS agent, the Terray extraction pipeline and the Lila protocol generator are all agents wrapped around a domain model: they plan, call tools, read output and decide what to do next, and the planning layer is usually a general model rather than a specialist. That layer is where a single key across more than 200 models at provider list price with 0% markup earns its keep, and where automatic failover matters — an agentic run that takes six hours is a run you do not want to lose at hour five because one upstream provider had a bad afternoon. The routing DSL composes those calls; the specialist stays where it belongs.

Which one to pick

If you are a computational chemistry or molecular simulation group with GPU capacity, Nemotron for Science is close to a no-brainer: permissive licence, a 50,000-question chemistry set you can both evaluate and train on, and a documented tool loop that has been demonstrated on the exact failure mode you deal with weekly. The absence of a vendor rate card is not a problem when you were going to self-host anyway.

If you are doing translational biology and need a model that reasons about proteins and disease mechanisms, Rosalind is the only one of the two aimed at your work, and the trusted-access review is the price of admission. Budget for the review timeline in your project plan, and note that the academic programme's free seats do not include API access — a university group that wants a pipeline needs its own commercial arrangement.

And if you are building the agent and not the model, the choice is less consequential than it looks. Both are callable from a loop; both have failure modes; neither publishes the completion rate that would settle the question. The sensible move is to build the loop so the model underneath it is a configuration value, then measure. That is the only comparison either vendor has left to you.

A screenshot of the Artificial Analysis model page for Nemotron 3.5 Super VL, showing the model title, the NVIDIA open-weights label with an October 2026 release date, and the performance and price analysis panel with Intelligence Index 14, 291 output tokens per second, $0.20 input and $0.50 output per million tokens, a 75% cache discount, 11.85 seconds to first token and $0.56 per task, above the site's written analysis and a technical specifications sidebar.