DeepSeek-R1 Token Counter
Estimated token count and cost for DeepSeek-R1, worked out in your browser.
A reasoning model, and the cost shape follows
R1 thinks before it answers. Those reasoning tokens are billed as output, and at $2.18 per million against $0.50 for input, output is 4.4 times the price here. That is a very different bargain from DeepSeek-V3, where output costs barely more than input.
So the two models in the same family want opposite habits. On V3 you can let it write. On R1 the thinking is where the money goes, and you cannot see it in the reply.
Why this page cannot predict your bill
Reasoning tokens are generated by the model, not supplied by you, so nothing you paste determines how many there will be. A hard problem produces many; an easy one produces few. Two prompts of identical length can differ severalfold in cost.
The input count on this page is still useful for checking you fit in the window and for the input half of the bill. For the output half, run a representative sample through the API and read the reported usage. Estimating it from the visible answer length will understate it badly.
More detail
The count is an estimate
No browser-runnable tokeniser is published for R1, so o200k_base stands in. Reasonable on English prose, weaker on code and on Chinese.
128K context
The same window as V3, and the reasoning tokens consume part of it as they are generated. On a genuinely hard problem with a long chain of thought, the effective room for your input is less than the number suggests. Leave margin.
Questions
Why is my R1 bill higher than the token count implies?
Reasoning tokens, billed as output at $2.18 per million. They are generated before the answer, they are not visible in the reply, and no input-side count can predict them.
How is this different from DeepSeek-V3 on cost?
V3 charges roughly 1.5 times more for output than input, so verbosity is cheap. R1 charges 4.4 times more and adds invisible reasoning tokens on top, so the same habits cost very differently.
Do reasoning tokens use up the context window?
Yes, as they are generated. On a hard problem with a long chain of thought, the room actually available for your input is smaller than 128,000, so leave margin.
When is a reasoning model not worth it?
For formatting, extraction, classification and simple rewriting. There is nothing to deliberate about, so you pay for thinking that adds nothing and wait longer for the answer.