NVIDIA announced Nemotron 3.5 Super VL this week — an open, customisable multimodal model built for enterprise agents, combining reasoning and tool use with text, image, document and video understanding. It is a genuinely good release. It is also, by the time you read the announcement, already sitting in eight places in our catalogue at eight different prices.
What the catalogue actually shows
These are live rows, not a survey. Input price per 1K tokens, cheapest first:
- •DeepInfra — $0.0001 in / $0.0004 out
- •OpenRouter, Nebius, Vercel, AskCodi — $0.00015 in / $0.0005 out
- •Hugging Face — $0.0002 in / $0.0005 out
- •Baseten — $0.004 in / $0.013 out
- •OpenRouter free variant — $0, at a reported 1M context
Top to bottom, that is a factor of forty on input and roughly thirty-three on output, for weights that are byte-for-byte identical. Nobody is being dishonest. Providers differ on hardware, batching, quantisation, how much headroom they keep, and how badly they want the traffic. The model is the same; the business around it is not.
The practical consequence is that “how much does Nemotron Super VL cost” has no answer. It has eight answers, and picking one by hand on launch day means committing to whichever you read about first.
Why a cheap multimodal model matters more than a cheap text model
Vision pricing has historically been where agent budgets quietly go. A screenshot is not one token, it is hundreds to thousands depending on how the provider tiles it, and a browser agent that screenshots every step pays that on every step. A text-only loop that costs cents can become a multimodal loop that costs pounds without anything in your code looking different.
So a competent VL model at $0.0001 per 1K input is a different proposition from a competent text model at the same price. It changes which loops are affordable to run continuously rather than on demand — document triage, chart reading, a UI agent that actually looks at the screen instead of guessing from the DOM.
One caveat worth stating because it will bite someone: capability metadata differs between our rows for this model. Some providers report vision and tool-calling explicitly and some do not, because we infer those flags per row and the inference lags a launch by days. If you are pinning a specific provider for a vision workload, send one real image through it before you build on the assumption.
Using it
Pin a provider if you have a reason to — a compliance boundary, an existing contract, a latency measurement you trust:
curl https://antbase.ai/v1/chat/completions \
-H "Authorization: Bearer $ANTBASE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/nemotron-3.5-super-vl",
"messages": [
{"role": "user", "content": "What is the trend in this chart?"}
]
}'Or do not pin it, and let the router treat the eight rows as what they are: one capability available at eight prices with eight availability profiles. That is the case where a routing layer earns its place — not by being clever about which model is best, but by noticing that the same model is forty times dearer on one provider than another and that the dear one is sometimes the only one up.
The free row, and what it is for
OpenRouter lists a free variant at a reported million-token context. Free tiers are rate-limited and their availability moves, so this is for evaluation rather than production — but for deciding whether Super VL reads your charts well enough to bother integrating, free and rate-limited is exactly the right shape.
ant:free will route to it when it is up and to something else when it is not, which is the only sane way to consume a free tier.
