After the launch of GPT-5.6, the self-cost is reduced by 20%.

According to TrendWatch Monitoring, OpenAI has revealed that after the launch of GPT-5.6 Sol, it has already begun participating in improving its production system. It uses Codex to analyze real traffic, adjust request allocation, and autonomously rewrite GPU kernels. The related optimizations have reduced the end-to-end running cost of the model by 20%.

GPT-5.6 Sol has also enhanced its accompanying draft model. It autonomously designs and runs hundreds of architecture experiments, initiates and monitors training. In cases of hardware failures or unstable training, it also intervenes. Ultimately, the efficiency of token generation through speculative decoding (small models first predict, main model batch validates) has increased by over 15%.

OpenAI describes this process as a continuous feedback loop: observe the production environment, identify bottlenecks, modify the system, and then validate the overall effect. GPT-5.6 has been involved in multiple stages of this process.