How GPT-5.6 Combines State-of-the-Art Intelligence and Efficiency

OpenAI designed the GPT-5.6 model family to balance performance and cost according to diverse task needs. This family maximizes efficiency not only through model intelligence but also through optimization across Inference and the Agentic Harness.

The model family consists of the following:

• GPT-5.6 Sol: The top-tier reasoning model, delivering better performance than Claude Fable 5 at less than half the cost.
• Terra: Maintains GPT-5.5-level intelligence while cutting cost in half.
• Luna: The fastest and cheapest model, cutting cost by 80% compared to Sol.

The core of this update is maximizing intelligence-per-token efficiency. Beyond simply becoming smarter, the models are trained to take more direct paths when performing tasks, accomplishing more with fewer tokens.

Additionally, the following technical optimizations were applied for infrastructure efficiency:

• Inference optimization: Generates more tokens on the same hardware through load balancing, Speculative Decoding, caching, and kernel optimization.
• Agentic Harness improvements: Increased agent efficiency through context bloat management, and optimization of tool use and repetitive tasks.

In particular, GPT-5.6 Sol played a key role in improving overall system performance, including autonomously rewriting and optimizing production kernels within Codex.