Where do your LLM tokens actually go? We measured ours

When we looked at what our agent was sending, most of it wasn't the question. It was search results, tool output and old messages the model didn't need for that step.

On our demo agent (5 questions):

• Before: 52.6K tokens per run, $0.129

• After trimming: 9.1K tokens, $0.057

• 83% fewer tokens sent to the model

Nothing about the model or the prompt changed. We only stopped sending context that didn't matter for the task.

That's what Contextwaise does: an AI relevance model scores every piece of context against the current task and drops what your LLM doesn't need, in about 100 ms.

Have you checked what your app really sends per call? Which part was the biggest surprise: tool output, RAG chunks or chat history?