OpenAI Launches GPT-5.6 Family: Flagship Model Sol Beats Claude Fable 5 on Coding at Half the Cost

OpenAI launches its GPT-5.6 model family, with flagship model Sol outperforming Claude Fable 5 on coding benchmarks at less than half the cost, while autonomously optimizing its own inference stack to cut serving costs by 20% and boost token-generation efficiency by over 15%.

• OpenAI's GPT-5.6 model family delivers a new balance of capability and cost, with flagship model Sol outperforming Claude Fable 5 on coding benchmarks at less than half the cost, while lighter models Terra and Luna offer further price reductions down the intelligence curve.
• GPT-5.6 Sol autonomously optimizes its own inference stack by rewriting production GPU kernels, improving load balancing, and enhancing speculative decoding, collectively reducing end-to-end serving costs by 20% and boosting token-generation efficiency by over 15%.
• The agentic harness powering Codex and ChatGPT Work cuts costs further by preventing context bloat through deferred tool discovery, capping tool output at 10,000 tokens, and preserving append-only prompt prefixes to maximize prompt-cache hit rates across multi-step agent tasks.