What did OpenAI’s Monday announcement reveal about the launch of its GPT-5.6 model family—Sol, Terra, and Luna—inside AWS’s Kiro spec-driven development environment, including how Kiro turns product intent into requirements, technical designs, and task lists; the joint Terminal-B

OpenAI said GPT‑5.6 Sol, Terra, and Luna are now available in AWS’s Kiro at kiro.dev, positioning the integration as a lower-cost, more structured way to use coding agents in production workflows. 7

• How Kiro works: Developers begin with product intent; Kiro converts it into a spec consisting of requirements, a technical design, and sequenced implementation tasks. The resulting artifacts provide persistent, reviewable context rather than relying solely on a one-off prompt. 457

• Benchmark claim: OpenAI and AWS said joint testing on Terminal-Bench 2.1 found that Terra completed successful tasks in Kiro at about 82% lower cost. This is an end-to-end cost-per-successful-task result, not an 82% accuracy gain. The reported task-success scores were 87.4% for Terra and 88.8% for Sol—so Sol was modestly more accurate in that comparison. OpenAI and AWS did not publish a decomposition separating savings attributable to Kiro’s agent harness from those attributable to Terra’s token efficiency or pricing. 7

• AWS’s rationale: The pitch is that spec-driven context makes coding agents more reliable and governable for teams: requirements, design choices, and tasks become an auditable source of truth, helping agents build against intended behavior rather than improvising from a short prompt. 347

• Model pricing context: On August 21, OpenAI reduced Sol’s list price from $5 to $4 per million input tokens and from $30 to $20 per million output tokens—a promotion stated to last at least through November 21, 2026. 1 The reduction applies to pay-as-you-go API use and token-credit pricing for ChatGPT Work and Codex; it does not increase the included usage limits of ChatGPT Pro, Plus, or Business plans. 26

• Earlier reductions: On July 30, OpenAI cut Luna’s price by 80% and Terra’s by 20%. 8 Reuters characterized the Sol reduction as occurring amid greater pricing and capability pressure from Anthropic and Chinese model providers. 2

• AWS relationship: The Kiro launch is also evidence of a deepening OpenAI–AWS commercial relationship. Reported milestones include an initial $38 billion compute agreement in November 2025, a February 2026 expansion to $100 billion over eight years, Amazon’s planned $50 billion OpenAI investment, and OpenAI’s commitment to consume roughly 2 GW of AWS Trainium capacity. The retrieved evidence supports the $50 billion investment and 2-GW Trainium commitment, but I did not find a primary source in this search result set independently substantiating every term and date of the $38 billion-to-$100 billion sequence. 910

The practical takeaway is that Terra is being marketed as the cost-efficient default inside Kiro, while Sol remains the slightly higher-scoring option; the headline 82% figure should be read as a combined system-level cost result, not proof that Terra itself is 82% more efficient in every deployment.