Kiro: About 82% Cost Reduction with GPT-5.6 Terra on Terminal-Bench 2.1

Tests on Terminal-Bench 2.1 report that with Kiro, GPT-5.6 Terra performs tasks while reducing costs by approximately 82%. OpenAI and AWS claim to have optimized the environment together and plan to continue this work. The GPT-5.6 model family, including Sol, Terra, and Luna, is available in Kiro for long and contextualized developments.

Tests indicate approximately 82% lower costs with Terra

OpenAI and AWS state that they have collaborated to optimize the Kiro environment as well as the OpenAI models. On Terminal-Bench 2.1, tests report that GPT-5.6 Terra successfully completed tasks in Kiro, achieving a cost reduction of about 82%. The Kiro setup, based on specifications, anchors the model from the outset in requirements, technical designs, and the context of the task. According to its description, this approach allows for quicker functional solutions with fewer errors. OpenAI and AWS announce their intention to continue improving the performance of OpenAI models in Kiro and to assist teams in extracting more value from AI throughout the development cycle.

AWS and OpenAI emphasize flexibility and cost control

Swami Sivasubramanian, Vice President of Agentic AI at AWS, states that the goal is to provide developers with the latest base models and expand their options to accelerate native AI development in Kiro. He also expresses excitement about adding the GPT-5.6 family to support complex and long-duration tasks. At OpenAI, Colleen Kapase, Vice President of Global Strategic Partnerships and Ecosystems, asserts that integrating GPT-5.6 into Kiro offers greater flexibility to adjust intelligence, speed, and cost at every stage of the software lifecycle. She adds that with AWS, OpenAI helps teams optimize every dollar invested while speeding up processes when time is of the essence.

Specifications, context, and controls structure development in Kiro

Kiro transforms a general intention into precise specifications, technical schematics, and tasks ready to be executed. This structuring of context is touted as simplifying GPT-5.6's understanding of objectives, expected system behaviors, and desired outcomes. In this framework, teams can turn product ideas into implementation plans, handle complex multi-step coding tasks, and enforce specification-driven development. Work is conducted within the context of the codebase and team standards, with checkpoints to review and refine the model's proposals. The correction of implementations can be verified through property-based testing.

Availability of GPT-5.6 and functional scope in Kiro

The GPT-5.6 family is available in Kiro and includes the Sol, Terra, and Luna models in the planning, building, reviewing, and testing stages. These models are presented as capable of helping produce higher-quality code with fewer iterations, while improving cost-effectiveness per token. GPT-5.6 is described as delivering more useful work per token, with better performance per dollar and on-demand capacity for complex tasks. Within Kiro, these features are utilized for long-term projects based on requirements, the codebase, and team standards. According to the presentation made to developers, this would result in a greater amount of work accomplished, a reduction in unnecessary efforts, and an optimization of cost-effectiveness in each session.