Kiro makes GPT-5.6 Terra 82% cheaper

OpenAI and AWS say GPT-5.6 Terra completed Terminal-Bench 2.1 tasks in Kiro at roughly 82% lower cost per successful task, according to the launch report. The result is a workflow claim, not a standalone model-price claim: Kiro supplies structured requirements, design, task context and property-based checks before the model writes code. Coding teams should file the number beside the archive’s case for model-harness economics: the cheapest successful agent may be the one given better specification and verification scaffolding, not simply the smallest model.