Anthropic Says Claude AI Surpasses Human Experts in Complex Physics Calculations

Anthropic has reported a significant development in artificial intelligence-assisted scientific research after its Claude model helped physicists extend a complex theoretical calculation from eight loops to nine. The work involves mathematical calculations used to describe how subatomic particles interact, an area of physics where even small increases in complexity can require substantial amounts of time and specialised expertise. The achievement was highlighted in October 2026 as an example of how advanced AI systems are beginning to contribute to difficult research problems rather than simply answering questions or generating computer code. According to reporting by IBM, the calculation represents a meaningful step in testing whether AI models can reliably support scientists working on problems at the frontiers of theoretical physics.

The calculation concerns a concept known as a scattering amplitude, a mathematical quantity physicists use to predict the likelihood of particles interacting in particular ways. In quantum field theory, researchers often organise complicated calculations into contributions called loops, which account for increasingly complex interactions and corrections. Each additional loop can dramatically increase the mathematical difficulty of the problem. Moving from eight loops to nine is therefore not simply a matter of adding one more straightforward step; it requires handling more complicated expressions and ensuring that the resulting mathematics remains consistent. Such calculations can help scientists test theoretical predictions and improve their understanding of fundamental particle interactions.

The result is notable because it suggests that AI can assist with research tasks requiring sophisticated mathematical reasoning. Earlier AI applications in science often focused on summarising published studies, searching scientific literature or helping researchers write and debug code. The latest work demonstrates a more demanding use case: helping experts advance a calculation that sits close to the limits of established theoretical methods. Physicist Anastasia Volovich of Brown University told IBM that the result was significant because it could help establish greater confidence in the ability of advanced AI systems to perform difficult calculations reliably. However, the finding should be understood as evidence of AI-assisted scientific progress, not proof that AI has generally surpassed human physicists across every area of research.

Anthropic’s broader strategy has increasingly focused on developing Claude models capable of carrying out complex technical and professional work. Its Claude 5.5 family includes models designed for different workloads, ranging from high-level reasoning and software development to faster, lower-cost tasks. The company has also promoted the use of AI agents that can interact with tools and support multistep workflows. In scientific settings, such capabilities could help researchers explore mathematical approaches, check intermediate results, automate repetitive calculations and investigate alternative solutions. Nevertheless, scientists must still verify AI-generated work, since sophisticated models can produce plausible-looking errors or make assumptions that are not justified by the underlying mathematics.

The development comes as technology companies compete to demonstrate that advanced AI can contribute to genuine scientific discoveries. Researchers are investigating potential applications across mathematics, physics, chemistry, biology and materials science, where progress can depend on analysing large datasets and solving problems that are difficult to approach through conventional methods alone. If AI systems can reliably assist with complex calculations, they could reduce some of the time required to test theoretical ideas and allow researchers to investigate more possibilities. However, the value of such systems will ultimately depend on reproducibility, independent verification and whether other researchers can confirm the results. The reported nine-loop calculation is an encouraging example of AI supporting high-level scientific work, but it is one step in the longer process of determining where these systems can be trusted to contribute meaningfully to research.