OpenAI’s GPT-5.6 family is now available inside Kiro, the AWS-backed coding platform built around specification-driven development. It is the first time Kiro has offered OpenAI models alongside its existing Anthropic options.
The release covers Kiro’s IDE, command-line interface and web app. Developers can choose between three versions of GPT-5.6 depending on the job. Sol is the flagship for difficult, long-running work such as large refactors and complex terminal tasks. Terra aims for a middle ground between capability and price, while Luna is designed for faster, high-volume work at the lowest cost.
All three models come with a 272,000-token context window. Kiro charges different credit multipliers for each tier: 2.4x for Sol, 1.2x for Terra and 0.6x for Luna. That makes the model picker less of a popularity contest and more of a practical budget decision. A team can reserve Sol for the genuinely hard problems, use Terra for day-to-day development and send repetitive work to Luna.
The more interesting part of the announcement is how the models are being used. Kiro turns a product request into requirements, a technical design and a set of executable tasks before the coding agent starts changing files. That structure gives the model checkpoints to work against instead of relying on one long prompt and hoping it stays on course.
OpenAI and AWS say GPT-5.6 Terra completed successful tasks on Terminal-Bench 2.1 at roughly 82% lower cost when used in Kiro. Kiro also reports strong benchmark results for Sol, including 88.8% on Terminal-Bench 2.1. Those are company-reported tests, so they should not be treated as a guarantee for every codebase. Still, the cost-per-completed-task framing is more useful than a raw benchmark score because it measures whether an agent actually finishes the job.
GPT-5.6 support is rolling out experimentally to Kiro Pro, Pro+, Pro Max and Power customers in AWS’s Northern Virginia and Frankfurt regions, with cross-region inference available. Users may need to restart the IDE or CLI, or refresh Kiro on the web, before the new models appear.
Sources: OpenAI and Kiro.
