If you build your own harness, you can drive that cost down even further.

Out-of-the-box agent harnesses are, in my opinion, way too bloated. They try to solve way too many things that aren't useful for everyone.

You are paying a huge premium for no good reason.

While cost is not the only reason to build your own harness, it's one of the biggest benefits.

A few optimizations that make it possible:

- optimize for tasks you care about
- lighter and more optimal system prompt
- more efficient tool calling
- better context compaction/handoff
- leverage different models via custom routing
- improved reliability through custom verifiers
... and much more.

Just some quick thoughts on it, but I really think it's important for folks reading the OP to understand that token cost is not the only thing to consider in all of this.