At 25 Cents A Task, How Does Nemotron 3.5 Super VL Perform?

🔍 Read the full analysis: At 25 Cents A Task, How Does Nemotron 3.5 Super VL Perform? 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, 2026, with Artificial Analysis reporting an Intelligence Index score of 26.5 and an estimated $0.25 per index task. The open-weight model is fast and available for commercial use under its stated license, but competing models in the supplied comparison score higher for similar or lower task costs, making its appeal depend on speed, deployment control and licensing needs.

Artificial Analysis places Super VL first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. That ranking is specific to the class: across open-weight models of all sizes, the source says it ranks about 18th. The model has 120 billion total parameters and 12 billion active in a mixture-of-experts design, supports a reported one-million-token context window, and accepts text and images while producing text.

The listed API rates are $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million. Artificial Analysis estimates a full index task costs $0.25, compared with $1.64 for Nemotron 3 Super. The source attributes the lower task bill partly to reduced output: roughly 46,000 output tokens per index task, versus 114,000 for the predecessor. It says output accounts for less than three cents of the new model’s task cost, while most of the bill comes from input and cache writes.

Speed depends on where the model runs. On a pre-release DeepInfra endpoint using NVFP4 weights, it produced about 630 output tokens per second with 100,000-token prompts. Across six providers, however, the reported median output speed was 196 tokens per second on standard prompts and 307 with 100,000-token prompts, with a wide provider range. Median time to first token was 0.89 seconds, but the source says the model takes about 11 seconds to reason before its first answer token; a 500-token answer took about 13.6 seconds end to end.

Speed Versus Benchmark Value

The figures frame a practical choice for teams buying model access or running workloads on their own hardware. Super VL’s low estimated task cost and strong throughput can suit batch jobs, long-context processing and background agents. But the benchmark comparisons in the source suggest it is not the strongest option if the priority is intelligence score per dollar.

The source lists Xiaomi’s MiMo-V2.6-Pro at 46 points for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and GLM-5.3-Flash at 42 for $0.28. Inkling Small is listed at the same score as Super VL, 26, for $0.09. Those comparisons make clear that the $0.25 figure is not, by itself, evidence of best value. Buyers also need to weigh latency, endpoint variation, license terms, hardware requirements and workload-specific accuracy.

For organizations that need a permissively licensed US open-weight model or want the option to self-host on NVIDIA hardware, those considerations may outweigh a lower benchmark score. The model’s reported improvement in knowledge-work and hallucination tests could also matter to particular applications, though benchmark performance does not guarantee results in a company’s own production tasks.

What the Benchmarks Measure

Artificial Analysis released its numbers on the model’s launch day, according to the source. Its Intelligence Index is an aggregate measure, and the reported 26.5 score should be read alongside the specific task results rather than as a complete measure of operational performance. The source’s “first” ranking applies only to open-weight models in the stated parameter class; it does not mean Super VL leads every open model or every use case.

The comparison with the prior model is more favorable: Nemotron 3 Super scored 13 and cost an estimated $1.64 per index task. Super VL’s score is roughly double, while the estimated task cost is substantially lower. The source also reports GDPval-AA, a benchmark for economically useful work, rising to 1,162 Elo from 490. Its AA-Briefcase score is 999 Elo, compared with 876 for Nemotron 3 Ultra.

Some reported results point to both progress and remaining limits. AutomationBench-AA is listed at 14.8%, against 3.0% for Nemotron 3 Ultra, while the absolute score remains modest. On AA-Omniscience, the reported hallucination rate fell from 87% for the predecessor to 36% for Super VL. These are benchmark figures in the supplied source, not guarantees that the model will achieve the same error rate in other settings.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Real-World Fit

The reported performance figures do not establish how Super VL will behave across production workloads. The 630-token-per-second result came from one pre-release DeepInfra endpoint, while the six-provider median was much lower; the source does not identify enough detail here to explain every provider’s hardware and serving configuration. The actual speed and cost for a particular customer will depend on its provider, prompt lengths, caching and output patterns.

The supplied material also does not provide independent testing beyond the cited Artificial Analysis results, nor does it specify how the benchmarks map to an individual organization’s tasks. The source describes the OpenMDW-1.1 license as allowing commercial use, but buyers should review the license text and any provider-specific terms before deployment. The model is released in NVFP4, and the exact hardware and operating costs of self-hosting are not detailed.

Finally, the article’s cost-per-task figure is an estimate tied to an Artificial Analysis index workload. It should not be treated as a universal price for an application, where long prompts, repeated context, caching behavior and agent retries can change the bill.

Test the Intended Deployment

For teams considering Super VL, the next step is to test the model on representative prompts through the provider or hardware they plan to use. Record latency, throughput, input and output usage, cache costs, and error rates rather than relying only on launch benchmarks. A direct comparison with alternatives should use the same tasks and evaluation criteria.

Further provider measurements and independent evaluations could clarify whether the high single-endpoint throughput is reproducible and how results vary by workload. Until those comparisons are available, the reported evidence supports a qualified conclusion: Nemotron 3.5 Super VL offers a potentially useful mix of speed, lower estimated task cost and deployment flexibility, but its value depends on the operator’s requirements rather than its class ranking alone.

Key Questions

How much does Nemotron 3.5 Super VL cost per task?

Artificial Analysis estimates $0.25 per Intelligence Index task. That is a benchmark-specific estimate, not a fixed price for every application or workload.

What does its Intelligence Index score mean?

The source reports a score of 26.5. It ranks first among the 67 open-weight models in the stated 40-billion-to-150-billion-parameter class, but about 18th among open-weight models overall, according to the supplied material.

Is Super VL faster than other models?

It reached about 630 output tokens per second on one pre-release DeepInfra endpoint with 100,000-token prompts. Across six providers, the reported median was 307 tokens per second for that prompt length, so results depend on the endpoint.

Can businesses use the model commercially?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Organizations should check the license terms and the terms of their chosen hosting provider before deployment.

Does a lower task cost make it the best-value model?

Not necessarily. The supplied comparisons list several models with higher Intelligence Index scores for similar or lower estimated task costs. Super VL’s case may instead rest on speed, its license, NVIDIA-hardware deployment and self-hosting options.

Source: ThorstenMeyerAI.com