Multi-model agents can cost 11x more for the same result

In tests for the agent Chimera, using more models did not always improve the result. On a 12-task set, one mid-level model got 100% while using 846 tokens. A fusion setup, where several models answer, a judge chooses, and another model combines the result, also got 100% but used 9,526 tokens.

That was about 11 times more tokens for the same score. Fusion was therefore moved behind a -style cascade: try a cheap model first, check the answer with a free gate, move to a mid-level model, and use fusion only when the gate fails. That cascade reached roughly mid-level quality while using about one-twelfth of the fusion tokens.

A manager-worker setup also depended heavily on the task: for small documents it used 47% more tokens than a single agent, but for large multi-step document work it used 66.5% fewer tokens at the same 100% quality.

Key points

  • One mid-level model scored 100% on 12 tasks with 846 tokens.
  • The fusion setup also scored 100% but used 9,526 tokens, about 11 times more.
  • A -style cascade keeps fusion as a instead of the first step.
  • For small documents, used 47% more tokens than a single agent.
  • For large multi-step document work, scoped workers cut tokens by 66.5% at the same quality.
Read original