Grok 4 Token Counter
Estimated token count and cost for Grok 4, worked out in your browser.
Priced against the mid-tier, sized above it
Grok 4 is $3 per million input and $15 per million output, which is exactly Claude Sonnet's pricing. Where it differs is the context window: 256,000 tokens against Sonnet's 200,000.
That extra 56,000 tokens is not a headline feature but it is occasionally the deciding one. A codebase or a document that overflows 200K by a little fits here without restructuring anything.
The count is an estimate
xAI publishes no browser-runnable tokeniser, so this uses o200k_base as a proxy. Reasonable on English prose, looser on code and on languages without spaces. Treat it as close rather than exact, and leave headroom against the 256K limit.
More detail
The 5 to 1 output ratio
Output costs five times input, which is the standard shape and means conversational bills are dominated by generation rather than by prompt size.
The lever is the same as everywhere else in that shape: cap the maximum output length, ask for structured data instead of prose, and stop the model restating your input back to you. Trimming the prompt is the instinct and usually the smaller win.
Where the real comparison lies
At identical pricing to Sonnet, the choice is not about cost per token. It comes down to which model is better at your specific task and how much the extra context is worth to you. Count the same prompt on both pages, then run a real sample through each and compare the outputs.
Questions
How does this compare to Claude Sonnet on price?
Identically. Both are $3 per million input and $15 per million output. Grok 4 offers a larger window at 256K against 200K, so the decision comes down to task fit and context needs rather than cost.
Can I rely on this count for Grok?
No. It is estimated with o200k_base because xAI publishes no local tokeniser. Expect a few percent of drift on prose and more on code.
How do I reduce the cost of a conversational workload?
Attack the output. At a 5 to 1 ratio, generation dominates. Cap the response length, request structured output, and avoid prompting patterns that make the model repeat your input.
Does my prompt get sent anywhere?
No. The count is computed in your browser from a bundled vocabulary, so nothing you paste is transmitted.