OpenAI and AWS have optimized the Kiro development environment for GPT-5.6 models and achieved approximately 82 percent cost reduction for successful task completion with GPT-5.6 Terra according to their own tests. The optimizations were documented on August 24, 2026 and complement the price cuts announced on July 30, 2026 of 80 percent for GPT-5.6 Luna and 20 percent for GPT-5.6 Terra.

Since July 14, 2026, all three model variants – Sol, Terra, and Luna – have been available via the Kiro platform and integrated into IDE, CLI, and web interface. The 82 percent cost reduction was measured on Terminal-Bench 2.1, a benchmark for coding tasks.

Three Model Variants for Different Requirements

The GPT-5.6 family consists of three tiers with different performance-cost profiles:

GPT-5.6 Sol is designed for complex multi-step work – specification-driven implementation, long-term refactoring, and complex terminal tasks. The model achieves a score of 80 on the Coding Agent Index.

GPT-5.6 Terra positions OpenAI as a balanced solution for everyday development work between the flagship Sol model and the fastest Luna model.

GPT-5.6 Luna is the fastest and most affordable model in the family. According to OpenAI, Luna achieves nearly 99 percent lower costs for professional tasks (measured by Agents' Last Exam) compared to Fable 5. Luna runs approximately nine times faster than comparable previous models and can use tools as well as complete multi-step workflows. Performance is comparable to frontier models from a year ago, but at approximately 6 cents per dollar per task.

Three Components for Efficiency Gains

OpenAI attributes the cost reduction to three factors: First, GPT-5.6 models follow a more direct path through workflows. Second, optimized inference systems generate tokens more efficiently and keep hardware productive through better routing. Third, smarter context management in the Agentic Harness helps avoid redundant processing.

Fast Mode and Strategic Model Selection

On July 30, 2026, OpenAI introduced Fast Mode in the API, replacing the previous Priority Processing. For GPT-5.6 Sol, Fast Mode achieves up to 2.5-fold speed compared to standard processing at double the price. Model quality remains unchanged. Requests with Priority tag automatically use Fast Mode.

OpenAI recommends a results-oriented approach: organizations should define requirements such as stakes, error costs, urgency, and scale, then evaluate where additional intelligence brings measurable improvements. A practical example is a code workflow that uses Sol for uncertainty resolution and planning, then employs Luna for well-specified changes, testing, and evaluation.

Deployment at Enterprise Partners

Companies using the GPT-5.6 family include Replit, Notion, Ramp, Blitzy, Cognition, and Dust. OpenAI justifies the price cuts and optimizations with the mission to make advanced intelligence more accessible and affordable.