Llama 4 Maverick Token Counter
Estimated token count and cost for Llama 4 Maverick, worked out in your browser.
A note on how this is counted
Llama's own tokeniser is not what runs here. This page estimates using the Claude vocabulary as a stand-in, which is a different family from the o200k_base proxy used for most other estimated models on this site.
That makes it a proxy of a proxy, and the honest position is that the number is indicative rather than precise. It is fine for deciding whether a prompt is roughly 5,000 or roughly 50,000 tokens. It is not fine for sizing against a hard limit with no headroom.
Open weights change what the number means
Maverick can be self-hosted, and that alters the whole calculation. The $0.30 and $0.85 per million quoted here are what hosted providers charge. Run it yourself and there is no per-token price at all: you pay for GPUs by the hour whether they are busy or idle.
Which is cheaper depends entirely on utilisation. At low or bursty volume, hosted per-token billing wins comfortably. At sustained high volume, dedicated hardware overtakes it. Token counts are how you find where your own crossover point sits, so count a representative workload and multiply.
More detail
A million-token window at a fraction of the price
Maverick's context window is a million tokens, matching Gemini Pro, at roughly a seventh of the input cost. Filling it once is about 30 cents.
The caveat is that hosted providers do not all serve the full advertised window, and those that do may price long context differently. Check what your provider actually offers before designing around the headline figure.
Not counted here
System prompt, tool definitions and images. Tool schemas are billed on every request, and on an agent workload they can outweigh the message itself.
Questions
Why is Llama counted with the Claude tokeniser?
Because no browser-runnable Llama tokeniser is available here, and the Claude vocabulary is the closest stand-in on hand. It makes the figure indicative rather than precise, so leave headroom when sizing against a limit.
Is it cheaper to self-host?
It depends on utilisation. Self-hosting replaces per-token billing with per-hour hardware, which wins at sustained high volume and loses at low or bursty volume. Count a representative workload and compare against your hardware cost.
Can I really use the full million-token window?
That is the model's capability, but hosted providers vary in how much they actually serve and how they price it. Confirm with your provider rather than assuming the headline number.
Does the text leave my browser?
No. The encoding runs locally in the page with a bundled vocabulary.