OpenAI has introduced the GPT‑5.6 model family, designed to balance advanced capabilities with cost-effectiveness. The flagship model, GPT‑5.6 Sol, reportedly outperforms Claude Fable 5 on coding benchmarks at less than half the cost. The Terra model matches GPT‑5.5 performance at a reduced price, while Luna offers the fastest and most affordable option, priced 80% lower than Sol.

These efficiencies are achieved through significant optimizations across the models, inference processes, and the agentic harness used by Codex and Chat GPT Work. With over 1 billion active users and 2 million businesses, OpenAI prioritizes distributing AI benefits widely. GPT‑5.6 represents their greatest intelligence-per-token efficiency yet, trained to perform more work with each token by taking a more direct path through tasks.

Key advancements include optimizing inference via load balancing, speculative decoding, and kernel optimization, enabling more output from the same hardware. The agentic harness is improved by better managing context, tool usage, and reducing repeated work. GPT‑5.6 Sol played a crucial role in autonomously landing many of these gains, accelerating inference and streamlining repeated tasks within the agentic harness.