OpenAI uses GPT-5.6 Sol to cut costs by up to 20%

OpenAI says it's using the GPT-5.6 Sol model to improve its GPT-5.6 series. This cuts end-to-end service costs by up to 20%. The model learns Triton and Gluon programming languages to rewrite and optimize GPU kernels in production. It also uses Codex to analyze traffic and fix load issues. In the forward pass, GPT-5.6 Sol finds work that can be precomputed or done in parallel, boosting GPU use. OpenAI uses the open-source tool FpSan to check the model's kernel code. It also improves speculative decoding, which makes token generation over 15% faster.