GPT-5.6 Sol Ultra 测试：局部推理显著提升，全局连贯性进步缓慢

What stands out from my tests with GPT-5.6 Sol Ultra is how good it is at concise elegant constructions and finding ways to expand and enhance existing arguments.

It feels different from GPT-5.5 Pro which excelled at finding counterexamples.

A shared weakness of both models is writing longer papers over longer human-AI interactions. It usually becomes a mess: either by plastering new results on top of the previous ones making a paper into a brain dump or cutting too many details so that the essence is gone.

However the coherence limit has moved. Previously anything over 20-pages long would automatically contain too many errors to be useful. Right now the model can write coherent 30-40 pages of mathematics - at least when it comes to some domains/arguments and over multiple prompts. But when it comes to heavy analytical computations and bookkeeping, the limit is still around 12-15 pages.

So in a way GPT-5.6 Sol has gotten significantly smarter at local-scale thinking, but its progress on global-scale thinking is much slower.