DeepSeek V4.1 Flash showing 108M tokens in ZCode traced to missing context compaction
Using DeepSeek V4.1 Flash in ZCode for small CSS fixes, the author saw 108 million tokens after 14–16 messages because the agent resends full history and never compacts; the same model feels normal in Claude Code.
If you've ever looked at your token counter and thought DeepSeek V4.1 Flash is eating tokens like crazy, here's what's actually happening. I've been using V4.1 Flash in ZCode for small CSS fixes, and after just 14 - 16 messages the counter showed 108 million tokens! but it turns out every message I send, the agent resends the whole conversation with it. Message 10 carries everything from 1 to 9, and one message can turn into dozens of requests behind the scenes. Most of that is cached and costs way less than it looks. The problem is that the harness never compacted my context, so the history just kept growing. Same model in Claude Code feels totally normal. so if your token counter looks insane, it's probably the agent, not the model.