Kernel of truth: GPT-5.6 Sol can cut its own costs, says OpenAI

OpenAI unveiled GPT-5.6 as a three-model family spanning premium, mid-tier, and low-cost options.

Its flagship model, GPT-5.6 Sol, reportedly beat Claude Fable 5 on a coding benchmark while generating 54% fewer output tokens.

The efficiency gains came from optimizations across four layers: model training, inference, the API stack, and the agent harness used by Codex and ChatGPT Work.

OpenAI also says GPT-5.6 Sol autonomously rewrote production kernels, helping cut end-to-end serving costs by 20%.

The story highlights how frontier models are becoming both products and infrastructure tools that can help lower their own operating costs.