The Case For Nemotron 3.5 Super VL’s 25-Cent Tasks

🔍 Read the full analysis: The Case For Nemotron 3.5 Super VL’s 25-Cent Tasks on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, and Artificial Analysis reported a score of 26 on its Intelligence Index and a $0.25 cost per index task. The model is a 120-billion-parameter mixture of experts with 12 billion active parameters, and its OpenMDW-1.1 license allows commercial use. Those figures make it a candidate for teams prioritizing speed, cost and deployment options, but the score alone does not establish that it is the best-value model for a given workload.

Artificial Analysis lists Super VL at 26.5 points and places it first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. It ranks about 18th among open-weight models overall, according to the source. Its predecessor, Nemotron 3 Super, scored 13. The release includes NVFP4, a four-bit weight format, and a one-million-token context window. The listed API prices are $0.20 per million input tokens and $0.575 per million output tokens; cached input is $0.055 per million tokens.

Speed varies by measurement. On a pre-release DeepInfra endpoint, the model produced about 630 output tokens per second with 100,000-token prompts. Artificial Analysis’s six-provider figures are lower: median output speed was 196 tokens per second on its standard prompt and 307 with 100,000-token prompts, with a reported range of 106 to 534. Median time to the first token was 0.89 seconds, while the source says the first answer token took about 11 seconds because the model reasons before responding. A 500-token answer took about 13.6 seconds end to end.

The source puts the cost of one Intelligence Index task at $0.25, compared with $1.64 for Nemotron 3 Super. It attributes the difference partly to lower output volume: about 46,000 output tokens per index task versus 114,000 for the predecessor. Its cost breakdown also puts roughly 89% of the task bill on input, including 69% on cache writes, and less than three cents on output. These estimates reflect the benchmark task and its token profile; actual bills depend on workload and provider.

The Cost of Speed and Control

Super VL’s strongest production case may be fast inference and deployment flexibility, rather than maximum benchmark score per dollar. Teams can access it through providers or run released weights on their own NVIDIA hardware, and the source says its license permits commercial use. That can matter when infrastructure control, data handling or licensing terms shape a deployment decision.

Its benchmark cost does not make it the cheapest route to a given score. The source compares Super VL’s 26 points and $0.25 task cost with Inkling Small, which also scores 26 but costs $0.09 per task. It lists MiMo-V2.6-Pro at 46 points for $0.13, and Claude Haiku 5.5 at 43 for $0.21. These comparisons use the source’s reported Artificial Analysis figures; they do not show how each model performs on a particular company’s tasks. The practical question is whether Super VL’s speed, license and hosting options justify its cost for that workload.

Input-heavy agents deserve particular scrutiny. Because the source attributes most of the benchmark task cost to input and cache writes, a team repeatedly feeding long documents or code repositories may see costs driven by how often those materials are processed. Token prices alone will not predict that bill.

How the Benchmark Compares

The Intelligence Index is a benchmark score, not a direct measure of performance on every product task. Artificial Analysis places Super VL first in a defined size class, 40 billion to 150 billion total parameters, but its overall open-weight ranking is lower. Both statements can be true because the comparison groups differ.

The source reports gains over Nemotron 3 Super in several evaluations. On GDPval-AA, covering economically useful work, Super VL scored 1,162 Elo versus 490 for its predecessor; it also exceeded Nemotron 3 Ultra’s 1,013 and Inkling Small’s 1,124. Its AA-Briefcase score was 999 Elo, compared with 876 for Ultra. On AutomationBench-AA, the reported score was 14.8%, around five times Ultra’s 3.0%, though the source describes that absolute result as low.

On AA-Omniscience, the source says the reported hallucination rate fell from 87% for Nemotron 3 Super to 36% for Super VL. That is a benchmark-specific result, not a guarantee that the model will avoid errors in other settings. The source also notes that the model’s Index average conceals uneven performance across evaluations.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider and Workload Variability

The reported throughput of about 630 tokens per second came from one pre-release endpoint. Artificial Analysis’s median across six providers was substantially lower, and its reported range was wide. Production speed will depend on provider and setup; the source does not establish that the pre-release result will be available to every user.

The $0.25 task figure is tied to Artificial Analysis’s index workload, and the source’s input-heavy cost breakdown may not match other usage patterns. It is also unclear from the material how Super VL performs on individual customer workloads, how its quality changes across all task types, or what hardware and operating costs are required for self-hosting. The benchmark’s lower hallucination rate likewise does not establish an error rate for a specific deployment.

Evaluate Your Intended Endpoint

Teams considering the model can compare it against alternatives on their own tasks, using the provider and hardware they expect to run. That evaluation should track latency, throughput, input and cache volume, output quality and total cost, rather than relying on a single benchmark score or endpoint result.

The next useful evidence will be production measurements across providers and workloads, alongside further task-level evaluations. The source does not give a timeline for those results or identify a forthcoming benchmark milestone.

Key Questions

What did NVIDIA release?

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. The source describes it as a 120-billion-parameter mixture-of-experts model with 12 billion active parameters and a one-million-token context window.

How much does an Intelligence Index task cost?

Artificial Analysis’s reported estimate is $0.25 per index task. It is a benchmark-specific figure, and actual costs depend on the task’s token use and the provider.

Is it the fastest model available?

The source reports about 630 output tokens per second on one pre-release DeepInfra endpoint with 100,000-token prompts. Across six providers, Artificial Analysis reported a median of 307 tokens per second at that prompt length, so the single-endpoint result does not establish typical speed across services.

Does it outperform other models for the price?

Not on the source’s reported Intelligence Index cost comparisons. It lists models with higher scores at similar or lower task costs, and Inkling Small at the same score for less. Speed, commercial licensing and the option to self-host may still affect which model fits a team’s needs.

What remains unknown about production use?

The source does not establish performance on specific customer workloads, actual self-hosting costs, or whether the pre-release endpoint speed will hold in production. Teams would need measurements using their own tasks and intended provider or hardware.

Source: ThorstenMeyerAI.com