In the latest offline IQ test by Tracking AI, multiple versions of OpenAI's GPT-5.6 all soared to 136 points, marking the first time a large language model has surpassed the 130 IQ threshold. In human IQ distribution, 130 is the starting line for "genius," with only about 1% of the global population reaching that level. This means GPT-5.6 is smarter than 99% of humans.

image.png

Scored 136 on the most difficult cheating-proof question bank, leaving all competitors far behind

Tracking AI has two sets of questions: one is an open-source Mensa Norway-style test, which models have already scored over 140; the other is an undisclosed, anti-leakage "offline question bank" designed to block models from memorizing answers beforehand. GPT-5.6 broke through this most challenging offline question bank. The entire SOL and TERRA family scored 136, even the visual version didn't fall behind. Following closely was Claude-5 Fable with 130 points, while GPT-5.6 LUNA Max and Claude-4.8 Opus were still hovering between 117 and 123 points. Over the past year, models from o3 to various flagship models kept getting stuck at the 130 threshold, but GPT-5.6 was the first to break through the door.

Not only does it do well in tests, but it also performs impressively in real tasks.

Just having a high score isn't enough. Developers tested GPT-5.6 in real work scenarios, and the results were impressive. Developer Amir Bohlooli fed the same physical simulation prompt to Fable5 and GPT-5.6 Sol, expecting to be overwhelmed by Fable, but instead, GPT-5.6 chose a particle fluid simulation, with physics progressing in real time. CSS, interface, and rendering were all packed into a single HTML file and automatically hosted as a shareable web page, generating a complete product with just one sentence. Ramanpal Singh also created a customer ticket system based on RAG with a single prompt, including four roles, a management backend, automatic complaint classification, and emotion recognition, building five applications at a cost only a fraction of Fable5's.

Claire Vo's experience was more vivid—she got stuck on a bug, thinking her code had crashed. After switching to GPT-5.6 Sol, she simply said, "I won't let this beat me." Sol fixed the issue in one go and even helped other models run smoothly. Her evaluation was straightforward: Fable's obsession with technical precision ultimately trapped it, while Sol's practical approach actually got the job done.