Gemini 3 Flash Token Counter
Estimated token count and cost for Gemini 3 Flash, worked out in your browser.
The cheapest way to hold a million tokens
Flash pairs the same million-token window as Gemini Pro with a quarter of the input price: $0.50 per million in, $3 per million out. If your problem is volume of context rather than difficulty of reasoning, that combination is hard to beat.
The count on this page is an estimate. Google publishes no browser-runnable tokeniser, so o200k_base stands in. Close on English prose, looser on code and on languages without spaces.
What the price makes possible
At $0.50 per million, filling the entire context window costs 50 cents. That changes which designs are reasonable. Feeding a whole document on every request instead of building a retrieval pipeline becomes defensible, because the engineering cost of retrieval exceeds the token cost at this rate for a lot of workloads.
The same arithmetic makes high-volume classification and extraction viable. A million short calls that would be uncomfortable on a frontier model are routine here.
More detail
Where it will not carry you
Flash is optimised for throughput, not for hard reasoning. On multi-step problems, subtle code, or anything where a wrong answer is expensive, the cheaper model is not the cheaper choice once you count the retries and the errors that get through.
The honest test is to run a representative sample through both Flash and Pro and compare outputs, not prices. If Flash is right often enough for your tolerance, the six-times cost difference is real money.
Not counted here
Images, audio and video, all of which Gemini bills by its own rules. The system prompt and any tool definitions are also excluded, and both are charged on every request.
Questions
How much cheaper is Flash than Pro?
Four times cheaper on input and four times on output, at $0.50 and $3 per million against $2 and $12, with the same million-token context window.
Is the token count exact?
No. It is estimated with o200k_base because Google publishes no local tokeniser. Expect a few percent on English prose and more drift on code and on CJK text.
When should I pay for Pro instead?
When the task involves multi-step reasoning or a wrong answer is costly. Test both on a real sample and compare accuracy, not price; if Flash is accurate enough, the cost difference is substantial.
Can I skip retrieval and just send the whole document?
Often, yes. At this price, sending a full document per request can cost less than building and maintaining a retrieval pipeline. Do the arithmetic for your own volume before assuming otherwise.