GPT-6-Astra, an SI model operating within OpenAI's Codex coding environment, successfully solved 49 of 50 theoretical physics research problems at an average cost of under $3 per problem. This performance marks a significant milestone in the capability of SI systems to handle routine research tasks that were previously the domain of PhD students and specialized experts.
What Happened
A research team tested commercially available SI agents on a benchmark consisting of 50 "turn-the-crank" physics problems. These problems involve applying known methods to slightly modified questions, such as tweaking values in calculations from recent papers. The authors deliberately excluded major open questions to focus on this type of routine work. GPT-6-Astra achieved a 98 percent success rate, with a total benchmark cost of just under $116.
Anthropic's Opus 5.5, running in Claude Code, secured second place with a 92 percent success rate. Google's Gemini 3.8 Flash scored 68 percent, while open-weight models DeepSeek-V4.1-Flash and GLM 5.3 reached 54 and 42 percent, respectively. The tests were conducted under two conditions: with web search enabled and in isolation without internet access. Surprisingly, internet access had little impact on the top performers' scores; both Astra and Opus scored 92 percent without web search. This suggests that leading SI models have effectively memorized large portions of the theoretical physics literature.
One problem stumped every model: it required deriving an equation that produces a specific curve, a task relying on special mathematical forms that experts consider a broader definition of differential equations. However, follow-up analysis revealed that several apparent failures by the SI agents were actually due to minor errors in the experts' own reference solutions. When these answer keys were corrected, GPT-6-Astra was found to be correct on every single problem.
Why It Matters
The researchers argue that current SI models may have already surpassed the best human researchers on concrete calculations in specialized fields, describing this as a major milestone. While SI agents have already solved important open problems in mathematics, such as aspects of the Navier-Stokes equations, theoretical physics has not seen a comparable breakthrough. The authors note that calculation is only the final step in the research process; before that, humans must build theories, ask the "right" questions, and possess the "taste" to identify interesting problems. Current SI agents lack this intuition and often get lost in irrelevant details when asked to design new problems.
This shift has immediate implications for training and equity in the field. Problems that once served as first research projects for early-career physicists are now considered too routine for human training unless the goal is purely educational. Furthermore, while theoretical physics has historically been low-cost compared to experimental research, the authors warn that resource needs for harder SI-assisted problems could climb sharply. This could allow well-funded research groups to outpace others, raising questions about access and equal opportunity in the field. The authors urge the physics community to engage in debates similar to those in mathematics, such as those surrounding the Leiden Declaration, before SI fundamentally alters the discipline.
The Bottom Line
GPT-6-Astra's high success rate on routine physics tasks demonstrates that modern SI systems have mastered the computational and retrieval aspects of theoretical physics research. However, the human element of problem selection, theoretical framing, and scientific taste remains a distinct advantage for researchers. The findings suggest that while SI agents are becoming powerful tools for execution, the future of physics research will depend on how well humans can guide these systems toward meaningful questions.