When AI Outsmarts the Experts: The Physics Test Humans Failed

According to The Decoder, a new benchmark called TheoPhysEval put top AI agents to the test using 50 unpublished theoretical physics calculations. OpenAI's GPT-6-Astra led the pack, correctly solving 49 of them at an average cost of $2.40 per problem, totaling $116 for the entire test.

Why Expert Oversight Almost Misjudged the AI

When the model appeared to miss answers during evaluation, researchers discovered that the specialists' own reference solutions contained errors. Once those human mistakes were corrected, the AI's answers proved accurate every time, revealing that the model had effectively memorized large parts of complex physics literature.

How the Models Ranked on Complex Tasks

Anthropic's Opus 5.5 secured second place with a 92% success rate, while Google's Gemini 3.8 Flash scored 68%. Weaker models struggled significantly more, with DeepSeek-V4.1-Flash and GLM 5.3 reaching 54% and 42% respectively, often copying outdated formulas blindly.

Vale a pena para a sua rotina?

For routine calculations and junior-level research tasks, these results show that top agents can match specialist output at a minimal financial cost. However, researchers point out that current models still lack scientific 'taste'—the ability to choose which problems are worth asking next.