The model then «autonomously rewrote and optimized our production kernels,» OpenAI writes in a report.
That means it found work that could be «precomputed, avoided, or parallelized,» leading to a reduced serving cost of 20%.
On speculative decoding, which lets a model produce several tokens in one pass and reduce expensive GPU time, GPT-5.6 Sol designed and ran hundreds of experiments and monitored «the speculator training process.» The end result was improving token efficiency by «more than 15%.»
Not only that, but Sol now runs continuously in Codex to analyze production workloads previously thought too large for manual intervention and «hyper-optimizes» how the engine behind GPT is configured «for each scenario,» optimizing «every part» of the inference loop.
This frees up the team to explore more ideas and makes for lower latency and fewer inference tokens for users. It is also some of the first signs of how AI can be used to atonomously improve on itself in the compute loop and return real-world results, although you can be sure that most modern models are already coded with the help of their own coding engines, as with Claude.
Read more: OpenAI’s report, X post. Discussion on r/Singularity.