[GPT-5.6 Launch Reduces Its Own Costs, Operating Costs Down by 20%] According to monitoring by Beating, OpenAI disclosed that after the launch of GPT-5.6 Sol, it has begun contributing to the optimization of the production system that hosts itself. By using Codex to analyze real traffic, adjust request allocation, and autonomously rewrite GPU kernels (the underlying code controlling chip computations), related optimizations have reduced end-to-end operating costs by 20%. GPT-5.6 Sol has also improved the accompanying draft model. It independently designed and conducted hundreds of architectural experiments, initiated and monitored training processes. When encountering hardware failures or unstable training, it intervenes to handle the issues. Ultimately, token generation efficiency in speculative decoding (where a smaller model predicts first and the main model validates in batches) has increased by over 15%. OpenAI describes this process as a continuous feedback loop: observing the production environment, identifying bottlenecks, modifying the system, and then verifying overall effectiveness. GPT-5.6 has already participated in multiple stages of this process. [Original Link]