When Anthropic released Claude 4 in January 2026, it came with a capability that had been rumoured for months but still managed to surprise the researchers who tested it: the model could spend hours working through a problem before producing an answer, revising its reasoning dozens of times in ways that looked, to human observers, uncomfortably close to genuine deliberation.
On the International Mathematical Olympiad — a competition that defeats the vast majority of professional mathematicians — Claude 4 solved all six problems correctly in its first full attempt. On a battery of PhD-level science questions spanning physics, chemistry, and biology, it outperformed the top 1 percent of human experts in the field.
“What we’re seeing is qualitatively different from what any model produced before,” said one researcher at Anthropic who has been evaluating the system. “It’s not just getting better answers. The reasoning chains it produces are long, coherent, and frequently novel. It finds approaches we didn’t expect.”
The extended thinking capability raises profound questions that the AI safety community is still working through. When a model thinks for three hours before answering, the internal representations it builds are extraordinarily difficult to interpret. The field of mechanistic interpretability — which tries to understand what is happening inside neural networks — is, for the moment, not equipped to audit that kind of reasoning.
Anthropic has released Claude 4 with significant capability restrictions for the most demanding reasoning tasks, requiring enterprise agreements and use-case review. The company says the caution is deliberate. Whether competitors will apply the same caution is an open question.
Leave a Reply