Counting tokens for Nemotron models
NVIDIA publishes the tokenizer with every Nemotron 3 checkpoint, so the counts on this page are exact. Nemotron 3 Nano and Super load by default. Add Ultra or Nemotron 3.5 Lightning from the compare table to price the same text on them; the count does not change, because every size ships the same vocabulary. Nano and Lightning show as free: with about 3B active parameters they run on your own hardware, so the context bar matters more than the cost.
Pricing and context windows
Nemotron input prices per million tokens, with the context window of each model.
| Model | Tokenizer | Price / 1M input | Context window |
|---|
| Nemotron 3 Nano 30B | nemotron-3 (tekken) | Free | 262K |
| Nemotron 3 Super 120B | nemotron-3 (tekken) | $0.09 | 1M |
| Nemotron 3 Ultra 550B | nemotron-3 (tekken) | $0.60 | 1M |
| Nemotron 3.5 Lightning 30B | nemotron-3 (tekken) | Free | 262K |
Prices and context windows are curated metadata refreshed periodically from the providers and OpenRouter, so treat them as guidance rather than a quote. Token counts are always produced by the real tokenizer.
How Nemotron tokenization works
Nemotron 3 uses Mistral's Tekken vocabulary, so it counts exactly like Mistral NeMo, Mistral 3, and Mistral 4. Comparing the files entry by entry, the 131,072-entry vocabulary and the merge rules in Nemotron 3 are those of Mistral's Tekken tokenizer. NVIDIA only renamed eight control slots to its own chat and reasoning markers, and those never appear in pasted text. Nano, Super, Ultra, Nano Omni, and Nemotron 3.5 Lightning share one file, as did Nemotron Nano 2 before them. On our fixed 500-word English sample the count is 581 tokens, the same as Mistral.
Tekken splits words with a case-aware rule and keeps each digit as its own token, which is why it runs a few tokens above OpenAI's encodings on English prose. Thinking is on by default in Nemotron 3, and the reasoning trace is billed as output, so a reply costs more tokens than its visible text. NVIDIA supports a thinking budget that caps that trace.
Frequently asked questions
Do all Nemotron 3 sizes use the same tokenizer?
Yes. Nano, Super, Ultra, and Nemotron 3.5 Lightning ship the same vocabulary file. Any text produces the same count on every size, so one counter covers the family. What separates them is the price, since Nano and Lightning are practical to run locally while Super and Ultra are hosted, and the context window a given host actually serves.
How large is Nemotron's context window?
NVIDIA quotes up to 1M tokens, but most hosts serve 256K. The model configuration files set 262,144 tokens, and reaching 1M needs explicit server flags and a lot of memory. The context bar here uses the window listed in the table, so check what your provider or local setup allows before planning a very long prompt.
Is this the same count as Mistral?
Yes. Nemotron 3 and Mistral share the Tekken vocabulary. Add a Mistral model from the compare table and the two counts match on any text. Only the special tokens for chat formatting differ, and those are not part of what you paste.
Are thinking tokens included?
No. This page counts the text you paste, which is input. Nemotron 3 thinks by default and writes its reasoning before the answer. That trace is output, billed at the output rate, and it also consumes context. You can turn thinking off in the chat template or cap it with a thinking budget.
