Can Nemotron 3.5 Super VL Keep Tasks Under 25 Cents?

🔍 Read the full analysis: Can Nemotron 3.5 Super VL Keep Tasks Under 25 Cents? on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, with Artificial Analysis reporting an Intelligence Index score of 26.5 and an estimated $0.25 per index task. The price is below the median reported for its model-size class, but comparisons in the source material show that several alternatives score higher for a similar or lower task cost, so the figure alone does not establish it as the best-value model.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a stated one-million-token context window. It accepts text and images and returns text. NVIDIA released its weights in NVFP4, and the source says its OpenMDW-1.1 license allows commercial use. API prices are listed at $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million tokens.

Artificial Analysis places Super VL first among 67 open-weight models in its 40B–150B parameter size class, according to the source. That is a size-class ranking, not an overall lead: counting open-weight models across sizes, the article puts it at about 18th. Its reported $0.25 task cost compares with $1.64 for its predecessor, Nemotron 3 Super, while its Intelligence Index score is reported as rising from 13 to 26.5.

Speed depends on the serving endpoint. A pre-release DeepInfra endpoint reportedly delivered about 630 output tokens per second on prompts of 100,000 tokens. Across six providers, the reported median was 307 tokens per second at that prompt length and 196 on the standard prompt; the source gives a 106-to-534 range across providers. Median time to first token was 0.89 seconds, but the source says the model reasons before answering, with the first answer token taking about 11 seconds and a 500-token response about 13.6 seconds end to end.

Cost Depends on the Workload

The quarter-dollar estimate gives teams a concrete way to compare benchmark performance with operating cost, but it is not a universal per-request price. Artificial Analysis calculates the task figure from its evaluation workload; real bills depend on how much input a system processes, how much it caches and how many output tokens it generates. In the source’s task-cost breakdown, about 89% of the $0.25 is input cost, with cache writes alone accounting for 69%; output costs less than three cents.

That makes Super VL potentially relevant for agents that repeatedly handle long documents or code repositories: cached input is priced at a discount, and the model can be run using released weights on NVIDIA hardware. But teams should test their own prompts and deployment setup. The source’s provider speed range, from 106 to 534 tokens per second, shows why a single endpoint result cannot predict every production experience. The model’s reported 11-second wait for its first answer token may also suit background workflows better than interactive chat.

Its value proposition is therefore broader than intelligence per dollar. The source compares it with MiMo-V2.6-Pro, which scores 46 at $0.13 per task, and Inkling Small, which scores 26 at $0.09. Those figures suggest Super VL is not the lowest-cost route to a given benchmark score. Its potential advantages instead include permissive commercial-use terms, NVIDIA deployment options, speed on some endpoints and its measured task profile.

How the Benchmark Compares

The Intelligence Index score is a benchmark measure, not a guarantee of performance on a particular company’s work. The source reports several areas where Super VL improved on Nemotron 3 Super. On GDPval-AA, a test of economically useful tasks, it rose to 1,162 Elo from 490; the source compares this with 1,013 for Nemotron 3 Ultra and 1,124 for Inkling Small. Its AA-Briefcase result was 999 Elo, compared with 876 for Nemotron 3 Ultra.

For AutomationBench-AA, the reported result was 14.8%, versus 3.0% for Nemotron 3 Ultra. The source also says its hallucination rate on AA-Omniscience fell from 87% to 36% compared with its predecessor. These are benchmark-specific findings; they do not show that the model will achieve the same results across every workflow.

The source’s price and score comparisons include MiMo-V2.6-Flash at 38 points for $0.06 per task, DeepSeek V4.1 Flash at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. It also lists Claude Haiku 5.5 at 43 for $0.21 and GPT-6 Luna at 38 for $0.07. These comparisons are reported figures in the source material, not independent testing in this article. Super VL is described as the top-scoring permissively licensed US open-weight model, while Meta’s Muse Spark Open scores 39 under a license the source says has commercial-use restrictions.

“Super VL is a very good engine and an average brain.”

— Thorsten Meyer, source article

Production Costs Still Vary

The reported $0.25 per task is tied to Artificial Analysis’ index evaluation, and the source does not establish that typical customer workloads will cost the same. Actual costs may vary with prompt length, caching, output volume, provider and hardware. The material also does not provide a full independent reproduction of the benchmark comparisons or specify how the scores translate to a particular business task.

Speed figures are similarly conditional: the 630-token-per-second result came from one pre-release endpoint, while measurements across providers showed a wide spread. The source identifies several benchmark strengths, but its excerpts do not include a complete account of all evaluation areas or failure modes. License terms and model weights are described in the source, but teams still need to check the applicable license and their own deployment requirements before relying on them.

Test the Intended Deployment

The next practical step for teams considering Super VL is to test it on their own workload and on the provider or NVIDIA hardware they expect to use. A useful evaluation should track input and cache usage, output volume, end-to-end latency and task accuracy, rather than relying only on per-token prices or the index score. That will show whether the task-cost estimate and speed results apply to their operating conditions.

Further provider measurements, independent evaluations and production reports may clarify how consistently the model performs across endpoints and use cases. Until then, the source supports a specific conclusion: Super VL combines a reported $0.25 benchmark-task cost with a 26.5 score, but whether that is a good deal depends on the workload, deployment choice and value placed on its licensing and hosting options.

Key Questions

Does Nemotron 3.5 Super VL cost less than 25 cents per task?

Artificial Analysis is reported to estimate a cost of $0.25 per Intelligence Index task. That is an evaluation-specific estimate, not a fixed charge for every user request; actual API bills depend on token use, caching and provider pricing.

Is it the highest-scoring open-weight model?

No. The source says it ranks first among 67 open-weight models in its 40B–150B parameter size class, but about 18th among open-weight models across sizes. Several listed alternatives have higher Intelligence Index scores.

How fast is the model?

A pre-release DeepInfra endpoint reportedly reached about 630 output tokens per second on 100,000-token prompts. The median across six providers was 307 tokens per second at that prompt length, and the reported first answer token took about 11 seconds because the model reasons before responding.

Can businesses use the model commercially?

The source describes the OpenMDW-1.1 license as allowing commercial use and says the weights are available for self-hosting. Organizations should review the license and deployment terms that apply to their intended use.

Source: ThorstenMeyerAI.com