GPT-5.6: Delivering frontier intelligence at unprecedented efficiency

OpenAI introduced the GPT-5.6 model family, engineered to balance advanced capabilities with cost-effectiveness across diverse tasks. The lineup features three main variants: GPT-5.6 Sol outperforms Claude Fable 5 on coding benchmarks at half the cost; GPT-5.6 Terra matches GPT-5.5's intelligence performance while cutting expenses in half; and Luna represents the most affordable option, priced 80% below Sol. The efficiency gains stem from optimizations across three system layers. The model itself was trained to accomplish more work per token. At the inference level—where models generate responses—the company optimized load balancing, speculative decoding, caching strategies, and kernel implementations, reducing operational costs by up to 20%. Speculative decoding improvements alone boosted token-generation efficiency by over 15%. The third component involves the agentic harness, a Rust-based orchestration layer managing interactions between models, tools, and user environments. Deferred tool discovery and context-window management minimize computational overhead, while prompt caching reuses previous computation. These measures drove high cache-hit rates in applications like Codex and ChatGPT Work. GPT-5.6 Sol played a key role in realizing these gains, autonomously analyzing production traffic, identifying bottlenecks, and rewriting GPU kernels for optimal performance. Together, these compounding improvements enable OpenAI to offer more capable models at lower costs simultaneously.