Token Snowball

Token snowball is the compounding growth in input costs as every new message in an AI session forces the model to reread the entire conversation history. What starts cheap quickly becomes expensive. The model does not store memory between turns. It reloads everything every single time. This is why a two hour design review can cost more to continue than it did to begin.

Token snowball is not just higher latency. It is not abstract compute talk. Most people treat it like a minor tax when it is the main reason long chats stop making financial sense. It is also not solved by bigger windows. Larger windows let the snowball grow bigger before it buries you.

Common confusion comes from looking only at output token prices. Input tokens are cheaper per unit but you pay them on every turn for the full history. A 80K token session means every new reply bills you for those 80K before it generates one new sentence.

Concrete example. A design team at Vercel in 2024 ran a single session reviewing a new dashboard with attached Figma files, previous iterations, and stakeholder notes. The first five turns cost under 8K input tokens each. By turn 25 with tool outputs and generated variants the same reply cost 62K input tokens. The snowball had taken over. They spent more rereading old decisions than creating new ones. A reset with a distilled one page spec dropped the next session to 9K input and restored speed.

Another case happened during Cursor coding sessions for a component library rebuild. Loading twelve files and test logs pushed context to 120K tokens early. Each subsequent edit cost 40 percent more than the last. The engineer finally started new sessions per feature. Token spend dropped 55 percent and bugs decreased because the model could actually focus.

Use token snowball awareness on any project that involves tools, file uploads, or more than 30 minutes of back and forth. Reset before 60 percent to keep costs predictable. Do not ignore it during quick explorations or single image generations. There the snowball never forms and a fresh chat adds unnecessary friction. The tradeoff sits in the middle. Continuity feels free until the snowball turns your clever workflow into an expensive reread exercise.

Smart operators track the percentage table like a fuel gauge. They prune dead branches at the healthy mark. They move finished decisions to external docs. The chat stays tactical. The snowball stays small.

The math is brutal but predictable. An 80K history session charges you for those 80K on turn 41, turn 42, and every turn after. Add image descriptions or long tool outputs and the curve gets steeper faster.

Teams that master this pattern build repeatable skills and prompt templates that live outside any single chat. Each new session inherits the system without inheriting the noise. Output quality rises. Bills shrink.

Token snowball proves that history is not free. Treat every turn like it carries the full weight of everything that came before. Because it does.

Keep sessions short and focused. The snowball stops before it starts.