Llama 4 Scout Token Counter
Estimated token count and cost for Llama 4 Scout, worked out in your browser.
Ten million tokens, with an asterisk
Scout's headline context window is 10,000,000 tokens, an order of magnitude beyond anything else on this list. It is a real architectural capability and it is also the number most likely to mislead you.
Almost no hosted provider serves the full window. Most cap it far lower, because the memory required to hold that much context is enormous and the attention cost at that length is punishing. Check what your provider actually offers before designing anything around ten million.
The second caveat is quality. Retrieval accuracy across extremely long contexts degrades regardless of the advertised limit. A fact buried in the middle of eight million tokens is not reliably found. Long context is not a substitute for retrieval; it is a different tool.
The cheapest tokens here
$0.15 per million input and $0.50 per million output, the lowest of any model on this site. At that rate a million-token prompt costs 15 cents, which makes designs viable that would be absurd elsewhere.
As with Maverick, that price is what hosted providers charge. Scout has open weights, so self-hosting swaps per-token billing for per-hour hardware, and which is cheaper comes down to how busy you keep it.
More detail
The count is doubly indirect
This estimates with the Claude vocabulary rather than Llama's own tokeniser. Combined with the very large window, that argues for generous headroom: an estimate that is a few percent off matters little at 5,000 tokens and matters a great deal when you are trying to land just under a limit.
Questions
Can I actually use ten million tokens of context?
Rarely. It is the model's architectural limit, but hosted providers almost always cap it well below that because of the memory and compute cost. Confirm the real limit with your provider before building on the headline number.
Does quality hold up across a very long context?
Not uniformly. Retrieval accuracy degrades as context grows, and a detail buried deep in a huge input is not reliably recalled. For precision, retrieval beats stuffing everything into the window.
How does the price compare?
It is the cheapest here, at $0.15 per million input and $0.50 output. A million-token prompt is 15 cents, which makes high-volume and long-context designs affordable that would not be on a frontier model.
How accurate is this token count?
Indicative. It uses the Claude vocabulary as a stand-in rather than Llama's own tokeniser, so leave meaningful headroom when sizing against any limit.