Field note
OpenAI's engineering account of GPT-5.6 attributes efficiency to a stack of changes rather than one model trick. GPT-5.6 Sol helped rewrite production kernels, improve load balancing, tune workload-specific engine configurations, and reduce end-to-end serving cost by 20 percent. A redesigned speculative-decoding draft model raised token-generation efficiency by more than 15 percent.
The agent harness was treated as part of the cost surface too. OpenAI moved more of it to Rust, deferred tool discovery, bounded tool output, and kept history append-only so repeated context could cache reliably. Those controls reduce the amount of text and repeated work that every later model call must carry.
The transferable lesson is architectural: token price is only one line in an agent budget. Routing, kernels, cache reuse, tool-output limits, and state design compound. Teams measuring only dollars per million tokens can miss the larger savings available in the execution path around the model.