People have long been debating the nature of consciousness. Is consciousness exclusively human, biological or can non-biological systems that are able to perform the right computations also be conscious? Since consciousness is potentially necessary for general intelligence, the answer to this question determines the future of artificial general intelligence as well. After all, if consciousness turns out to be non-computational, computers are fundamentally unable to be conscious and are therefore unlikely to achieve general intelligence. The current study aims to assess the ability to generalize in five frontier Large Reasoning Models (Gemini 3.1 pro, GPT-5.4, Claude Opus 4.6, Z.AI GLM-5 and KIMI K2.5). 

A novel game benchmark was developed. Complexity of the task was gradually increased to capture the ability to generalize to more complex environments. The results show that only Gemini was able to generalize to higher complexity levels, while the other models failed to generalize beyond the lowest complexity level. However, Gemini’s high latencies seem to indicate a brute force approach, rather than actual reasoning and generalization. All things considered, frontier LRMs do not seem to be able to generalize to novel, computationally infeasible situations as well as humans can, indicating a lack of general intelligence. This might be symptomatic of a more fundamental problem, namely the non-computational nature of consciousness.