A team of 35 researchers from Harvard, ICTS, KU Leuven, and Google DeepMind released TheoPhysEval, a new benchmark featuring 50 unpublished, verifiable problems in theoretical physics. The benchmark tests the computational ability of frontier AI models on standard research tasks that typically take a domain specialist up to a full day to complete.

Running in OpenAI’s Codex environment, GPT-6-Astra achieved a top score of 98 percent by correctly solving 49 out of 50 problems at an average cost of under $3 per problem. Anthropic’s Opus 5.5 scored second with 92 percent, while Google’s Gemini 3.8 Flash achieved 68 percent and open-weight models DeepSeek-V4.1-Flash and GLM 5.3 trailed at 54 and 42 percent respectively.

The findings show leading models can automate standard graduate-level computational workflows across 11 physics subfields. However, the study authors noted that breakthrough original research remains bottlenecked by the models’ lack of scientific “taste” and creative direction.

Why it matters

  • Proves frontier LLMs can automate routine, complex computational tasks across technical disciplines at minimal compute costs.

  • Shows top AI models have effectively memorized massive scientific literature, maintaining 92% benchmark accuracy even without internet access.

  • Highlights that domain breakthrough capacity remains constrained by high-level reasoning and problem formulation, not computational execution.

Source: the-decoder.com