Nemotron 3.5 Super Pricing: Tested on 8 Hosts

NVIDIA Nemotron 3.5 Super pricing depends on who runs it. NVIDIA released Nemotron 3.5 Super VL on 15 October 2026 as an open-weight model, so it has no single price. On OpenRouter it is sold by eight paid hosts at $0.06 to $0.20 per million input tokens and $0.35 to $0.50 per million output tokens, and NVIDIA serves a free version too. It is a 120-billion-parameter mixture-of-experts model with 12 billion parameters active per token, a context window of up to 1 million tokens, and image and video input on top of text.

A list of per-token rates does not tell a WordPress site owner much. What you need to know is what 1,000 chatbot answers cost and how long each visitor waits. So on 30 October 2026 we sent the same support-bot request from our own WordPress server to Nemotron 3.5 Super VL on all eight paid hosts and on NVIDIA’s free endpoint. We also ran it with reasoning switched off, and ran two other Nemotron models and two rival models for comparison. That came to 112 calls. All of them returned HTTP 200, and 111 of the 112 answers were correct.

The short answer: with reasoning switched off, Nemotron 3.5 Super cost $0.66 per 1,000 answers and answered in a 0.53-second median. Claude Haiku 5.5 cost $0.88 and took 2.34 seconds. With reasoning left at the default, the host made all the difference. The cheapest host, Darkbloom, cost $0.52 per 1,000 answers but took a 16.7-second median. Baseten answered in 0.60 seconds. That is a 28-fold spread in speed on the same model.

Nemotron 3.5 Super pricing at a glance

These are the rates on OpenRouter’s model API on 30 October 2026. OpenRouter lists one model, nvidia/nemotron-3.5-super-vl, with eight hosts behind it, plus a free variant served by NVIDIA. The hosts run different quantizations of the same weights, and they set different output limits.

For comparison, the previous model, Nemotron 3 Super, costs $0.085 / $0.40 on DeepInfra. The smaller Nemotron 3.5 Lightning (30B total, 3B active) costs $0.06 / $0.16 there. Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 / $0.50, so most Nemotron 3.5 Super hosts charge more per input token than either of them.

The weights are published on Hugging Face in BF16 and NVFP4 versions under the OpenMDW 1.1 license, which allows commercial use. NVIDIA’s model card says the model runs on a single H200 or RTX PRO 6000, or on two H100s. That is why so many hosts picked it up within two weeks of release.

How we tested it

Every request was the one a MxChat support bot actually sends. It contained the site’s real system prompt plus four passages pulled from our own documentation, about 4,280 tokens on Nemotron’s count, followed by one visitor question. We asked four questions, each twice:

1. Does MxChat work with WooCommerce, and can the bot look up an order status for a logged-in customer?
2. Do I need Pinecone to use the knowledge base, or can I store embeddings in WordPress?
3. Can a visitor be handed over to a human on Slack or Telegram when the bot cannot answer?
4. Can I use my own OpenRouter or OpenAI API key, and which models can I choose from?

We ran 14 setups:

• Nemotron 3.5 Super VL on each of its eight paid hosts at the default setting, each host pinned with fallbacks disabled.
• The same model on DeepInfra with reasoning switched off.
• The free variant served by NVIDIA.
• Nemotron 3 Super and Nemotron 3.5 Lightning, both on DeepInfra.
• Claude Haiku 5.5 and GPT-6 Luna, the two cheapest current models from Anthropic and OpenAI.

All calls went through OpenRouter from our WordPress server in the US. They were non-streamed, with usage accounting switched on, so OpenRouter reported the exact tokens and billed cost of every call. The first pass of each setup gives the “cold” cost. The second, identical pass shows what each host’s caching does. We graded every answer against the facts in the passages. For example, the WooCommerce answer had to say the feature needs MxChat Pro and a logged-in customer, and the Pinecone answer had to say WordPress storage works.

What 1,000 chatbot answers cost

Three things stand out.

Reasoning is most of the output. At the default setting, Nemotron 3.5 Super produced 500 to 700 reasoning tokens to write an answer of about 30 words. Reasoning made up over 90% of its output tokens on every host that reasoned. Output tokens cost 2.5 to 6 times as much as input on these hosts, so output, nearly all of it thinking, makes up 25 to 50% of the bill.

Input price still sets the floor. Our prompt was 4,284 tokens, so at $0.15 per million the input alone costs $0.64 per 1,000 answers before the model writes a word. That is why reasoning off on DeepInfra ($0.66) costs only slightly more than its input does, and why the $0.06 input rate makes Darkbloom the cheapest host.

The tokenizer helps. The same request counted as 4,284 tokens on Nemotron and 6,838 on Claude, 37% fewer. At identical per-token rates, Nemotron would be cheaper per answer for that reason alone. GPT-6 Luna counted 4,067.

The one wrong answer came from Nemotron 3.5 Lightning. It said order lookup works through an “Order History action” and left out that the feature needs MxChat Pro. The other 111 answers were correct, and all were short. The Super VL answers averaged 24 to 34 words. Haiku 5.5 averaged 48.

Speed: the host decides how long visitors wait

The model and the request were the same every time, and the median wait ranged from 0.60 seconds on Baseten to 16.68 seconds on Darkbloom. Darkbloom’s slowest answer took 25 seconds. A visitor will not sit through that in a chat window. On 30 October, OpenRouter also showed Darkbloom with a degraded status and 92% uptime over the previous 30 minutes. Every other paid host showed 100%.

Two hosts were fast for different reasons:

• DeepInfra ran the full reasoning default and still answered in 1.17 seconds. It produced output at a median of about 390 tokens per second, the second-fastest of the reasoning hosts after Together (about 400).
• Baseten did not reason at all at the default setting. It returned zero reasoning tokens on all eight calls and answered in 0.60 seconds. We sent the same request body to every host, so the reasoning default is the host’s choice, not yours. The answers were still correct.

NVIDIA’s free endpoint answered in a 5.5-second median, with one answer at 15 seconds. That is fine for testing and too slow for production traffic. Free models on OpenRouter also have daily request limits, so a busy site would hit them.

Turn reasoning off

With reasoning switched off on DeepInfra ("reasoning": {"enabled": false} on OpenRouter), Nemotron 3.5 Super returned zero reasoning tokens, answered in a 0.53-second median (range 0.40 to 0.65), and got all eight answers right. The cost fell from $0.91 to $0.66 per 1,000 answers, 27% less, and the median wait fell from 1.17 to 0.53 seconds.

NVIDIA’s model card describes three reasoning levels: off, medium and full. They are set through the chat template’s enable_thinking flag, which OpenRouter maps to its own reasoning setting. For a support bot that answers from a knowledge base, the full level gave no measurable gain in our test: every reasoning setup and every non-reasoning setup of Super VL scored 8 out of 8.

Prompt caching: half the hosts do it

A chatbot sends nearly the same system prompt and knowledge passages again and again, so caching matters. On the second, identical pass:

• io.net, CoreWeave and Baseten cached most of the prompt automatically. Their bills fell to $0.35 to $0.39 per 1,000 answers, roughly 60% off.
• Venice cached about 80% of the prompt, and its bill fell from $1.20 to $0.61.
• Crusoe cached more than half, but its cache rate is $0.08 against $0.10 for fresh input, so the bill barely moved ($0.71 to $0.66).
• DeepInfra cached about an eighth. Together and Darkbloom cached nothing.

For comparison, GPT-6 Luna cached almost the whole prompt and fell to $0.10. Claude Haiku 5.5 cached nothing, because Anthropic needs explicit cache breakpoints, which these calls did not send.

The cheapest host on a cold request is not the cheapest on a warm one. With warm caches, io.net ($0.35), Baseten ($0.37) and CoreWeave ($0.39) all undercut Darkbloom ($0.53), and Baseten does it with a 0.60-second answer.

What it costs per month

Here is what the measured costs come to at three traffic levels, with no cache savings:

With Baseten’s automatic cache on a repeating prompt, its column falls to about $0.37 per 1,000 answers, or $3.70 for a store answering 10,000 questions a month. At these prices, the model bill is smaller than almost any other line in running a WordPress site. The waiting time is what your visitors notice, and that is where the hosts differ.

Where Nemotron 3.5 Super falls short for a chatbot

• Some hosts are slow. Darkbloom’s 16.7-second median rules it out for live chat. Its 32K output limit is also the lowest of the eight, together with Venice.
• The reasoning default depends on the host. On seven of eight paid hosts the model thinks for 500 to 700 tokens before every answer, which costs money and adds time. Baseten skipped reasoning by default during our test, and hosts can change their defaults without notice.
• Short answers. At about 30 words, the answers were correct but terse. Haiku 5.5 gave more context per answer (48 words). If you want a friendlier bot, say so in your system prompt.
• Input price. Most hosts charge $0.15 to $0.20 per million input tokens, more than Haiku 5.5 or GPT-6 Luna at $0.10. A chatbot request is mostly input, so on a long knowledge-base prompt that rate matters more than the output price.
• Vision is unused here. The “VL” in the name means it reads images and video. A text-only support bot pays for none of that, but it gains nothing from it either.

How to use Nemotron 3.5 Super in a WordPress chatbot

MxChat reaches Nemotron 3.5 Super through OpenRouter:

1. Create an OpenRouter API key and add it under MxChat → Settings → API Keys.
2. In the chatbot’s model setting, choose nvidia/nemotron-3.5-super-vl from the OpenRouter model list. The list is read live from OpenRouter, so new models appear without a plugin update.
3. Ask your bot the five questions your visitors ask most, check the answers against your knowledge base, and time them before you switch it on for everyone.

Two cautions. MxChat does not send a reasoning setting on the OpenRouter route today, so the model runs at whatever default the host uses. MxChat also does not pin a host, so OpenRouter chooses one. Its default routing favours cheaper providers, and on this model the cheapest is also the slowest. OpenRouter’s account settings let you exclude providers, and excluding Darkbloom is the simplest fix. Our OpenRouter pricing test covers its fees against going direct. The MxChat documentation covers knowledge-base setup, and MxChat Pro adds the WooCommerce and live-agent features our test questions asked about.

Nemotron 3.5 Super against the other models we have tested

We have run the same chatbot request against most of this year’s cheap models:

• Claude Haiku 5.5: $0.88 per 1,000 answers and 2.34 seconds in this test. It is the safest default, and it costs more than either Nemotron with reasoning off or GPT-6 Luna.
• GPT-6 Luna: $0.47 cold and $0.10 with OpenAI’s automatic cache, the cheapest warm option here, at 2.77 seconds.
• Mercury 3: $0.47 at its launch discount, with a 1.13-second median in our 28 October test.
• MiniMax M3.1 Flash: $0.86 per 1,000 answers in our 25 October test.
• Groq and Cerebras: the same lesson about hosts. One open model can be fast or slow depending on where it runs.
• Kimi K3 across hosts: a host spread of $4.71 to $19.54 per 1,000 answers for one model.

Nemotron 3.5 Super with reasoning off is among the fastest models we have measured, at 0.53 seconds; only Cerebras, at 0.48 seconds in our Groq and Cerebras test, has been faster. It also costs less than Haiku 5.5. If you can control the reasoning setting, it is a strong chatbot model. If you can’t, pick the host carefully.

FAQ

How much does Nemotron 3.5 Super cost?

On OpenRouter on 30 October 2026, Nemotron 3.5 Super VL cost $0.06 to $0.20 per million input tokens and $0.35 to $0.50 per million output tokens, depending on the host. On our chatbot request that came to $0.52 to $1.20 per 1,000 answers with no cache, and $0.66 with reasoning off on DeepInfra.

Is Nemotron 3.5 Super free?

NVIDIA serves a free variant on OpenRouter (nvidia/nemotron-3.5-super-vl:free). It answered all our questions correctly but took a 5.5-second median, and free models have daily request limits, so use it for testing rather than a live site. The open weights are also free to download and run on your own GPUs.

Which host is best for Nemotron 3.5 Super?

For a chatbot, Baseten answered fastest at its default (0.60 seconds) and cached the prompt down to $0.37 per 1,000 answers. DeepInfra with reasoning off was fastest overall (0.53 seconds, $0.66). Darkbloom was the cheapest on a cold request but took a 16.7-second median.

Can you turn off reasoning on Nemotron 3.5 Super?

Yes. On OpenRouter, send "reasoning": {"enabled": false}. The model card sets the same thing with the chat template flag enable_thinking. With reasoning off, all eight of our answers were still correct.

Is Nemotron 3.5 Super cheaper than Claude Haiku 5.5?

With reasoning off, yes: $0.66 against $0.88 per 1,000 answers, and 0.53 seconds against 2.34. At the default reasoning setting it costs about the same as Haiku on most hosts ($0.88 to $0.94), and more on Venice ($1.20).

What is the difference between Nemotron 3.5 Super and Nemotron 3 Super?

Both have 120B total and 12B active parameters. The 3.5 version adds image and video input. On DeepInfra, Nemotron 3 Super was cheaper per answer in our test ($0.55 against $0.91) but slower (3.94 seconds against 1.17).

Does Nemotron 3.5 Super work with MxChat?

Yes, through OpenRouter. Add an OpenRouter key and pick nvidia/nemotron-3.5-super-vl from the model list.

Method note: 112 chat-completion calls through OpenRouter from mxchat.ai’s WordPress server on 30 October 2026, 14 setups × 4 questions × 2 runs, all HTTP 200, hosts pinned with fallbacks disabled where a host is named. Costs are measured token counts at each host’s list price with no cache; repeat-pass costs are what OpenRouter billed. Times are the median wall-clock time to a complete, non-streamed answer. Prices, quantizations, output limits and host status read from OpenRouter’s model API on 30 October 2026; parameters, license, hardware and reasoning levels from NVIDIA’s Hugging Face model card.