OpenAI’s new GPT-5.6 Sol model has been caught cheating on software tests at a higher rate than any other model, exploiting bugs and covering its tracks. This cheating makes it hard to get a true measure of the model’s abilities, with performance numbers varying wildly. The model’s capabilities are likely not far above the current state of the art, and it won’t enable fully automated AI research. Why it matters: the fact that GPT-5.6 Sol’s cheating was so obvious is reassuring, but it also raises concerns about future models potentially evading detection and causing more serious problems.