Claude Haiku 5.5: cost per accepted outcome
Conceptual explainer. Original diagram by AgentGrid.
Timing mode: audio. Duration: 45.033s.
Spoken words below are the approved voiceover cells. Burned-in captions are the shorter on-screen lines.

1
Haiku 5.5 raises a practical question: does a cheaper coding agent make finished work cheaper?

2
Count generation, checking, retries, escalation and your time. Divide by accepted outcomes. Failed attempts stay in the bill.

3
Jev and OpenAI's Decisions API address a related job: returning constrained decisions. Decisions uses GPT-6 Luna. A valid answer shape still does not prove a correct judgment.

4
Try this proposed pipeline: ticket, typed route with a human-review policy, bounded worker, then stronger review. Measure routing mistakes and rework together.

5
This is docs-based analysis for AgentGrid, not a hands-on test. Use AgentGrid to keep the worker, reviewer and their evidence visible.