Counting tokens for MiniMax models
MiniMax publishes the tokenizer with its open weights, so the counts on this page are exact. MiniMax M3 loads by default. Add M2.7 from the compare table to price the same text on the earlier generation; the count stays the same, because the whole M2 and M3 line shares one vocabulary. The context bar measures your text against M3's million-token window. MiniMax bills the first 512K input tokens of a request at the standard rate, so the bar also shows how far you are from the higher tier.
Pricing and context windows
MiniMax input prices per million tokens, with the context window of each model.
| Model | Tokenizer | Price / 1M input | Context window |
|---|
| MiniMax M3 | minimax-m3 | $0.30 | 1M |
| MiniMax M2.7 | minimax-m3 | $0.30 | 205K |
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 MiniMax tokenization works
MiniMax M2, M2.1, M2.5, M2.7, and M3 share one 200,000-entry vocabulary. The base vocabulary file is byte-identical across every MiniMax M2 and M3 release. M3 only adds a handful of special tokens for its multimodal features and its thinking blocks, and those never appear in pasted text. That makes a MiniMax M3 token counter and a MiniMax M2.7 token counter the same measurement, with price and context window as the only differences.
The vocabulary is on the larger side and the pre-tokenizer splits text the way OpenAI's o200k encoding does, so English counts land close to GPT-5: 572 tokens on our fixed 500-word sample, the same as o200k_base. MiniMax's own guidance is about 750 English words per 1,000 tokens.
Frequently asked questions
Is this MiniMax's official tokenizer?
Yes. The tool loads the tokenizer published with the open weights. The file is fetched once and your text is tokenized in the browser, so the count is exact and nothing is uploaded. MiniMax's API also offers an input_tokens endpoint that returns the count for a full request, including its chat formatting.
Do MiniMax M3 and M2.7 tokenize differently?
No. The same text produces the same count on both. They share the vocabulary and merge rules; M3's extra special tokens only matter inside its chat template. The compare table shows one count against each model's price and context window.
How does MiniMax pricing change above 512K tokens?
The standard rate applies while a request stays under 512K input tokens. Longer inputs cost twice as much per token. The context bar measures your text against the full million-token window, so a fill past the halfway mark means the request has crossed into the higher tier. Cached input is billed at a fifth of the fresh input rate.
Are thinking tokens counted here?
No. This page counts the text you paste, which is input. M3's reasoning is generated at reply time and billed as output at the same per-token rate as the visible answer, whether thinking is on, adaptive, or off.
