GPT-5.6 Sol participated in the optimization of its own infrastructure

GPT-5.6 Sol reduces OpenAI's operating cost by 20%. The model rewrote its own runtime components to optimize GPU computations.

The GPT-5.6 family does not rely solely on new models. OpenAI also describes a series of improvements made to its entire infrastructure in order to reduce the cost and time required for each response. This work involves query routing, GPU execution, caching, and tool orchestration in Codex and ChatGPT Work.

GPT-5.6 Sol was used in Codex to analyze production traffic, test new configurations, and rewrite certain components responsible for executing model operations. According to OpenAI, these optimizations reduced the overall operating cost by 20%. The company relies in particular on Triton and Gluon, two open-source languages designed for GPU programming, as well as verification tools to monitor the generated modifications.

The model also helped improve its own speculative decoding system. A lighter model proposes several tokens in advance, and then the main model verifies them in parallel. GPT-5.6 Sol reportedly conducted hundreds of experiments on this architecture and supervised its training, including during hardware failures or unstable phases. OpenAI measures a gain of over 15% in token generation efficiency.

Another part of the work focuses on Codex and ChatGPT Work. To prevent histories, tools, and integrations from unnecessarily saturating the context, functions are loaded only when they become necessary, and tool results are limited to 10,000 tokens by default. The history visible to the model also remains organized in a fixed order to more easily reuse previously performed computations through caching.

OpenAI thus positions GPT-5.6 Sol as a model capable of improving the systems that power its own operation, while Terra targets a balance between cost and performance, and Luna prioritizes speed. The company claims that Terra reaches a level comparable to GPT-5.5 for half the price, while Luna would cost 80% less than Sol.