OpenAI has revealed that its latest model, GPT-5.6 Sol, is already improving the production systems it runs on. Through Codex, it analyzes real traffic, adjusts request allocation, and autonomously rewrites GPU kernels (the low-level code that controls chip computations). These optimizations have cut the model's end-to-end operating costs by 20%.
GPT-5.6 Sol also improved its companion draft model. It designed and ran hundreds of architecture experiments, launching and monitoring training. When encountering hardware failures or training instability, it intervenes. The result: speculative decoding (where a small model predicts tokens and the main model batch-validates them) now sees token generation efficiency gains of over 15%.
OpenAI describes this process as a continuous feedback loop: observe the production environment, identify bottlenecks, modify the system, then verify overall results. GPT-5.6 is already involved in multiple stages.