# Nemotron 3.5 Super VL: Pricing and Benchmarks (2026) By Ingrid Kovarová · 2026-10-23 · Source: https://www.activepieces.com/blog/nemotron-35-super-vl-pricing-and-benchmarks-2026 --- If it misreads an image at 2am, the page goes to whoever owns the workflow it gets wired into (say, an automation built in [Activepieces](https://www.activepieces.com)), so the release details matter. Nemotron 3.5 Super VL is a new vision-language model from NVIDIA, meaning it **reads images alongside text**. ## Nemotron 3.5 Super VL release: what happened ### Who released Nemotron 3.5 Super VL NVIDIA is the maker, but the Hugging Face page is published by inference-snaps. It posted the model on its own Hugging Face page, the public hosting site where model makers publish weights and documentation. This article gives no release day, because a date taken from secondhand coverage would be a guess. The practical consequence is that you should treat NVIDIA's page, not a news recap, as the authority on what shipped. ![Activepieces connectors library page showing 502 available integration pieces with filtering options and sample connector…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/4edd1db5-aeb0-4a8b-ad9e-9b509394e8ac/power-automate-migration-guide-fixing-connector-4cf0cd0d.webp) ### Nemotron 3.5 Super VL Hugging Face page checked We read the page on 2026-10-23 for the publisher, what it says the model does, and how the maker describes access. It sits under the inference-snaps organization on Hugging Face. No third-party summary counts as evidence here. Knowing what exists does not tell you whether it fits your flow, and a model page only says what exists. Generic steps answer that question: a webhook receives a request and starts a flow, and an HTTP step sends a request to a model endpoint and returns the reply. Any automation tool's connector library lists these. Because those steps are generic, they apply to any model with an endpoint, and the rest of this integration uses them. ## Vision-language models for automation teams At 2am, the inputs that break automations are rarely clean text. They are scans, attachments and screenshots, and someone on call owns whatever the workflow does with them. A vision-language model matters because it reads images alongside text, which lets those items enter a workflow as data rather than as a manual queue. Four kinds recur: - Scanned invoices, feeding the step that pulls vendor, amount and due date into an approval record. - PDF attachments, feeding the step that classifies a document and routes it to the right owner. - Screenshots in support tickets, feeding the triage step that identifies the error and the product area. - Photographed forms, feeding the step that transcribes handwritten or skewed fields into a database entry. ![Scanned invoices stacked next to an approval record card](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/79e7fecb-2c29-43c7-9e5f-995bb523590b/nemotron-3-5-super-vl-pricing-and-benchmarks-202-af87c5ed.webp) On OCRBench v2, the benchmark [OCRBench v2](https://99franklin.github.io/ocrbench_v2/) uses to score document reading, Nemotron 3.5 Super VL scores **92.2 on extraction**. Pulling fields from invoices and forms is therefore the strongest case. Those are the maker-side scores; your own scans will settle the question. ## What's officially new in Nemotron The page states the model's size and architecture but gives no figure for access or speed, and those are the numbers an on-call owner will need. Only a tenth of Nano fires per token: [AIWiki](https://aiwiki.ai/wiki/nemotron_3) lists it at 30 billion total and 3 billion active parameters. Per AIWiki, Super is 120.6 billion total and 12.7 billion active, so you size memory for the whole model while compute tracks the smaller number. Ultra is 550 billion total and 55 billion active, which puts it beyond a small team's hardware. **Two rows are filled and three are empty**, so your comparison lacks sufficient data. The table below lists what the page states for this model, with the date of each check. | Item | Fact stated | Source | Date checked | |---|---|---|---| | Model size | 120B parameters | No page linked | 2026-10-23 | | Architecture | nemotron_h_moe | No page linked | 2026-10-23 | | Access method | not stated | not stated | 2026-10-23 | | Cost | not stated | not stated | 2026-10-23 | | Speed | not stated | not stated | 2026-10-23 | _Prices and plan limits checked against [huggingface.co](https://huggingface.co/inference-snaps/nemotron-3.5-super-vl) and [openrouter.ai](https://openrouter.ai/nvidia/nemotron-3.5-super-vl) on October 23, 2026._ You can't budget or set a latency limit from the launch material, because those rows are empty. **Measure both on your own scans before anyone owns this in production.** ## Nemotron integration in automation workflows Use a generic HTTP request step to call whatever endpoint serves the model, and move to an AI step later if that endpoint speaks the OpenAI API format. No dedicated integration for this model exists, so someone on your team owns the glue either way. ![HTTP request step sending a payload card to a server](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/2448b7e2-7b60-46fe-b874-8721720ef067/nemotron-3-5-super-vl-pricing-and-benchmarks-202-f4e96e98.webp) Read the table as a trade between effort and control; it compares the three generic routes. | Route | Setup effort | Control over the request | Fit for the task | |---|---|---|---| | HTTP request step | Low | Full: headers, body, image encoding | Good for a first test on your scans | | AI step, custom OpenAI-compatible provider | Low to moderate | Limited to what the provider exposes | Good once the prompt is stable | | MCP server | Highest | Set by the server you write | Worth it only if agents need the model as a tool | Start with the first two routes, which cost little, and keep the third for a real tool-calling need. Reselling you a model is deciding your AI strategy for you, and a release like Nemotron 3.5 Super VL shows the cost: a platform with its own bundled model can't offer you a new one on day one. [Activepieces](https://www.activepieces.com) has no Nemotron integration either, but its AI step takes a custom OpenAI-compatible provider, so you point it at whichever endpoint serves the model and the spend lands on your provider account. On the Plus plan, at $20 a month, bringing your own key makes every AI step 1 credit, meaning you can calculate your total monthly credit needs by counting your steps. MoneyGram, Moneypenny, Alan and FundingSocieties run Activepieces, and Claude, ChatGPT, Cursor or any MCP client can reach it. With an HTTP request step, you build the payload yourself, including the image encoding. That's more work per step, but a malformed request shows up in the logs, which helps the person on call. Only if your host exposes that API format does an AI step with an OpenAI-compatible provider work. Self-hosting changes the ownership question: the [Hugging Face](https://huggingface.co/inference-snaps/nemotron-3.5-super-vl) listing gives 120B parameters, so a single small GPU won't hold it, and you also own that hardware. ![Huge crate beside a tiny graphics card](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/d15d6b65-c932-4335-8f9a-9556fbda3255/nemotron-3-5-super-vl-pricing-and-benchmarks-202-866b22a5.webp) When agents must call the model themselves, an MCP server pays off. It wraps the model as a tool, and it adds a service to maintain. ## Nemotron 3.5 Super VL test results versus GPT-5.4-mini On our small test, Nemotron 3.5 Super VL was slower and costlier per call than openai/gpt-5.4-mini. The extra service only pays off if its answers on your documents are better. ### Nemotron 3.5 Super VL test tasks and method On 2026-10-23, both models got the same three automation tasks: sorting a support ticket, pulling fields from an invoice, and summarising an email thread. We called them through [OpenRouter](https://openrouter.ai/nvidia/nemotron-3.5-super-vl) with default settings, which means both reason before answering. This is three tasks, run once, by us, so we get one chance to execute correctly. It's indicative, not a benchmark, and it says nothing about your invoices or your ticket queue. ### Results, speed and cost per task Across the three tasks, the median Nemotron 3.5 Super VL call took 1.5 seconds and cost $0.000363, against 0.8 seconds and $0.000150 for openai/gpt-5.4-mini, which means Nemotron's competitive advantage, if it exists, cannot come from latency or cost. Individual calls will deviate from these medians. The Nemotron call **took nearly twice as long** and cost about 2.4 times as much. At 10,000 calls a day, that's roughly $3.63 against $1.50, which is small money. The wait matters more if a person is watching the screen. Accuracy comes first when a field comes back wrong at 2am, so rerun these three tasks on your own samples and compare the answers before the timings. ## Nemotron 3.5 Super VL cost and speed comparison Cost and speed come second to accuracy. The table below sets Nemotron against the baseline model from our test and the usual alternatives, and it only counts once your own samples agree. It shows how each option is reached and billed, and what our small test can and cannot say about speed and quality. | Option | Access method | Cost | Speed in our test | Result quality in our test | |---|---|---|---|---| | Nemotron 3.5 Super VL via OpenRouter | One API key and a model ID on OpenRouter, a routing service that fronts many providers | Billed per token at the rate on its OpenRouter page | One small run, so a sample rather than a benchmark | Usable on our three tasks, with no claim beyond them | | openai/gpt-5.4-mini (our baseline in this test; OpenAI's current efficient model is GPT-6 Luna) | Same OpenRouter key, different model ID | Billed per token at its listed rate | Our existing baseline, timed the same way | Our reference answers; the comparison point | | Other common options, such as small hosted vision models from Anthropic, Google or Mistral | Each vendor's own API, or the same OpenRouter key | Per-token pricing that differs by vendor | Not tested by us | Not tested by us | Because both models sit behind one key, swapping them costs you a changed model ID rather than a new integration. Whoever owns the workflow should treat our timings as a prompt to run their own, since provider routing and image size can shift speed. Cost is the one column you can settle from published pricing alone. ## What to track with Nemotron Watch the maker's own page and your own three documents, not the launch coverage, because only those tell you who owns the model and what it does for you. Because NVIDIA is the maker, its model card is the record of ownership. Copy the release date and version string from it into your runbook; whoever is paged at 2am needs to know which build is running. Check the model card and the repository for licence, input limits and version, and record the day you read them. **Prices move faster than pages**. [Dataku](https://dataku.ai/blog/cost-of-ai-dropped-97-percent-two-years-one-chart) tracked GPT-4-class output at $60 per million tokens in March 2023, $30 in November 2023, $15 in May 2024, $0.60 in July 2024, $0.40 in December 2024 and $0.30 in April 2025. ![Output price for GPT-4-class capability has collapsed](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/1e6c8de4-a712-46e3-8acd-e023f1a4dbe3/nemotron-3-5-super-vl-pricing-and-benchmarks-202-20931452.svg "Source: Dataku") , so any price you budget today is a snapshot to re-measure. One day is enough for the trial below, and it captures that snapshot: 1. Write the known answer for each before running anything. 2. Run both models through one HTTP step with identical input. 3. Record correctness, seconds and cost per task. 4. Compare the two sets of records. Whichever model wins on your three tasks, keep the sheet, since it becomes the baseline for the next re-test, which means future models must beat this benchmark to replace it. ## What Activepieces does about this The problem in this article is a gap between a model launch and a workflow you can trust, and Activepieces closes it at the step level. **There is no Nemotron integration**, and we won't pretend otherwise. If the endpoint doesn't speak that API format, the HTTP request step still lets you build the payload and the image encoding yourself. That fits the one-day trial above. Both models can sit behind the same flow with identical input, which means the comparison costs you a changed model ID rather than a new build. The spend lands on your own provider account, so the per-call figures from our small test (a median of $0.000363 for Nemotron against $0.000150 for openai/gpt-5.4-mini) are the numbers you'd see, not a marked-up credit price. Count the steps in your flow and you have the platform side of the bill. The model side you measure from your own scans, which is the point of the trial. The people who get paged also benefit from how Activepieces is built. Each run is logged step by step, so a malformed request or a misread invoice shows up as a specific step, not a vague failure. MoneyGram, Moneypenny, Alan and FundingSocieties run Activepieces. Claude, ChatGPT, Cursor or any MCP client can reach it, so if agents later need the model as a tool, that route is open without rebuilding the flow. ## Frequently asked questions What is Nemotron 3.5 Super VL? It's a vision-language model from NVIDIA, meaning it takes images alongside text and returns text. For automation teams, that covers work like reading a screenshot, a scanned form or a dashboard and producing something a workflow can act on. Does Activepieces support Nemotron 3.5 Super VL? We found no maker's page announcing a dedicated integration, so don't assume one exists. The practical route is the generic building blocks covered in the integration section, and the dedicated section on that platform covers what it offers. Whoever owns the workflow should confirm the current state before depending on it. Can I call Nemotron 3.5 Super VL from a workflow today? Yes, through generic steps such as an HTTP request to a hosted endpoint, provided you have credentials and a place to run or reach the model. Decide first who's paged when that endpoint fails, because a generic step gives you no vendor support path by default. How did it compare with GPT-5.4 mini in your test? Our test was small, and it doesn't support a ranking. Run your own tasks rather than reading this as a verdict. Different documents, image quality and output formats can reverse a result, so your own scoring sheet is the only comparison that should drive a decision. ## Related reading - [Self Host Nemotron 3.5 Bolt: Setup Guide (2026)](https://www.activepieces.com/blog/self-host-nemotron-35-bolt-setup-guide-2026) - [Best Data Automation Tools of 2026 for Analytics Teams](https://www.activepieces.com/blog/best-data-automation-tools-of-2026-for-analytics-teams) - [Reverse Logistics Automation Tools for Operations Teams](https://www.activepieces.com/blog/reverse-logistics-automation-tools-for-operations-teams) ## References - [OCRBench v2](https://99franklin.github.io/ocrbench_v2/) - [AIWiki](https://aiwiki.ai/wiki/nemotron_3) - [Dataku](https://dataku.ai/blog/cost-of-ai-dropped-97-percent-two-years-one-chart)