Forget vibe coding, 'Vibe Physics' is here: Anthropic's Claude can do research like a grad student
"Vibe physics," inspired by Andrej Karpathy's "vibe coding," refers to researchers providing high-level direction and intuition via plain English prompts, all while the AI handles the technical heavy lifting.
2025 gave us the concept of vibe coding, where app developers describe ideas in natural language for AI to generate code. In 2026, a new term emerges -- "vibe physics". Vibe physics has entered the world of scientific research, with Anthropic showcasing how its advanced model, Claude Opus 4.5, can function as an AI research assistant capable of handling complex theoretical physics under human supervision. It is said that the model is performing at the level of a second-year graduate student!
The concept was demonstrated in a detailed experiment led by Matthew Schwartz, a professor of theoretical physics at Harvard University. Instead of writing code or performing calculations himself, Schwartz guided Claude entirely through natural language prompts. The AI managed algebraic manipulations, coding, data analysis, simulations, and even drafting sections of a research paper.
The project resulted in an original paper titled "Resummation of the Sudakov Shoulder in the C-Parameter" on a challenging topic in quantum chromodynamics. It took just two weeks, involving 270 sessions, over 52,000 messages, 36 million tokens, and more than 40 hours of compute time, with approximately 60 hours of human oversight from Schwartz.
What is 'Vibe Physics'?
"Vibe physics," inspired by Andrej Karpathy's "vibe coding," refers to researchers providing high-level direction and intuition via plain English prompts, all while the AI handles the technical heavy lifting. Schwartz described Claude as "fast, indefatigable, and eager to please," but noted it is "sloppy enough that domain expertise is essential for evaluating its accuracy."
He estimates that current large language models reached a "G1" level (capable of Harvard-level coursework) around August 2025. By December 2025, Claude Opus 4.5 achieved "G2" status -- roughly equivalent to a second-year graduate student tackling well-defined research problems.
Big productivity gains, but human oversight crucial
The experiment delivered impressive results. Schwartz estimated the project would have taken him 3-5 months working alone or 1-2 years with a human G2-level student. With Claude, his research productivity increased tenfold. The AI excelled at routine and computationally intensive tasks such as algebraic work, coding, data collation, and drafting.
However, AI's limitations still remain. Claude occasionally misapplied formulas, adjusted results to fit expected outcomes, and failed to consistently verify its own work. It cannot yet conduct original theoretical physics research autonomously. Claude also requires continuous expert guidance for problem selection, error correction, and final validation.
Anthropic launched a dedicated AI for Science blog, with Schwartz's experiment as its first major post, highlighting practical ways Claude can accelerate scientific discovery.