OpenAI has used its latest artificial intelligence model, GPT-5.6, to directly optimise its AI infrastructure, lowering inference costs and improving processing efficiency.

Online outlet Gigazine reported on Wednesday that OpenAI improved its inference engine and GPU usage using GPT-5.6. It said this cut end-to-end service costs by 20 percent for its high-performance GPT-5.6 Sol model and improved token generation efficiency by 15 percent.

The case reflects operational efficiency gains from the GPT-5.6 series unveiled on July 9. The GPT-5.6 series consists of the high-performance Sol, the balanced Terra, and the low-cost Luna. Among them, GPT-5.6 Sol was introduced as delivering high performance at about half the cost of a rival model, Claude Fable 5.

The key is that GPT-5.6 Sol was directly used to optimise OpenAI's inference infrastructure. OpenAI explained it improved core kernels of its AI inference engine through GPT-5.6 Sol. The model learned kernel optimisation techniques based on the GPU programming languages Triton and Gluon, boosting code efficiency and cutting overall service operating costs by 20 percent.

It also improved a method to increase inference speed called Speculative Decoding. The technique has a small draft model predict the next token first, then a larger model verifies it. GPT-5.6 Sol was also used to improve the draft model's performance, and as a result token generation efficiency improved by 15 percent, the company said.

OpenAI also explained it optimised KV cache handling and its GPU task distribution system using GPT-5.6. It said it allocated computing resources more efficiently during inference so it could handle more requests on the same hardware.

It also revised how it runs agents. It changed MCP servers and various skills to be called only when needed and limited tool output to a maximum of 10,000 tokens by default. It designed the model to request additional output only when more is needed. It also fixed the order of tool calls to raise cache utilisation.

OpenAI explained that these improvements reduced unnecessary computation and variability in the agent execution process, while boosting both inference performance and cost efficiency.

OpenAI said, "GPT-5.6 contributed greatly to improving the inference engine and harness," adding, "We are confident the pace of optimisation will accelerate further."

In the industry, the case is seen as meaningful in that it shows development beyond competition over model performance to a stage where AI optimises AI infrastructure itself. Going forward, it is expected that not only model performance but also how much a system can improve service operating costs and GPU resource utilisation on its own will emerge as a new benchmark of AI competitiveness.