What's this about?
OpenAI pointed its GPT-5.6 Sol model at its own serving infrastructure. The model produced production GPU-kernel improvements worth ~20% lower serving cost and ~15%+ better token-generation efficiency. Separately, OpenAI says two harness/API setting changes let the same model reach state-of-the-art on ARC-AGI-3, a hard 2D-puzzle benchmark it had been failing.
What exactly is a "harness"?
The harness is everything wrapped around the model that isn't the model itself: how it's prompted, how much it's allowed to remember between steps, what tools it can call, how many tries it gets. OpenAI found Sol was failing ARC-AGI-3 not because it wasn't smart enough, but because the harness wasn't letting it remember what it had already learned. Flipping two settings tripled the score. The lesson people are drawing: a lot of "the model can't do X" is really "the harness won't let it."
What's the community saying?
- A model rewriting the kernels that run itself is the self-improvement loop people both hype and fear — here it's mundane efficiency, not runaway recursion.
- Skeptics note "20% cheaper" is an infra win that seasoned performance engineers also deliver; the news is that a model did it unattended.
- The ARC-AGI-3 story fuels the "it's the harness, not the model" camp (see Mollick, same day).