All three members of OpenAI’s GPT-5.6 family, Sol, Terra, and Luna, are now available inside Kiro, AWS’s spec-driven development environment, the companies announced Monday.
The integration covers the full Kiro workflow, from turning product requirements into structured implementation plans to executing multi-step coding tasks and checking results with property-based tests. Kiro’s pitch is context: it converts high-level intent into requirements documents, designs, and task lists before the model starts generating code.
That approach drove the headline number. In joint testing on Terminal-Bench 2.1, successful tasks completed with GPT-5.6 Terra came in at roughly 82 percent lower cost than the baseline, a vendor-run figure about cost rather than accuracy. The companies credit the spec-driven harness with cutting wasted tokens on missteps.
The pricing story has moved fast since the family’s July 9 launch, when Terra cost $2.50 per million input tokens and $15 per million output. A July 30 update cut Luna’s price by 80 percent and Terra’s by 20 percent, and OpenAI has since trimmed Sol’s API pricing too.
The release is another step in a deepening AWS-OpenAI relationship that began with a $38B compute deal last November and expanded by $100B in February, with Amazon investing $50B in OpenAI and committing the lab to roughly 2 gigawatts of Trainium capacity.
