GPT-6-Astra tops theoretical physics benchmark with 98 percent accuracy

Benchmark results

Thirty-five researchers from the International Centre for Theoretical Sciences, Harvard University, KU Leuven, and Google DeepMind released TheoPhysEval, a benchmark with 50 unpublished theoretical physics research problems with verifiable answers. GPT-6-Astra, running in OpenAI's Codex coding environment, answered 49 of 50 correctly, or 98 percent, at an average cost of under $3 per problem and just under $116 for the full benchmark.

Anthropic's Opus 5.5, running in Claude Code, scored 92 percent. Google's Gemini 3.8 Flash scored 68 percent, while DeepSeek-V4.1-Flash and GLM 5.3 reached 54 and 42 percent. Each agent ran twice, with web search and without internet access.

What the problems test

The benchmark covers eleven subfields, from fluid flows and biological processes to particle physics and cosmology. The authors focused on routine research work rather than major open questions, describing problems as turn-the-crank tasks in which known methods are applied to slightly changed values. A leading expert would need about a day or less on a typical problem, while neighboring specialists would take longer.

Internet access made little difference. Without web search, Astra and Opus each scored 92 percent. The authors say this suggests leading language models have effectively memorized large parts of the theoretical physics literature. Several misses were later traced to flaws in the reference solutions, not model errors.

From calculation to discovery

The authors say some of the strongest models may already surpass the best human researchers on concrete calculations, but theoretical physics has not seen a milestone like AI solving open problems in mathematics. They argue that verifiable calculation comes at the end of research. Before it, researchers must build a theory, ask the right questions, and switch when a problem is too hard. Current agents lack that judgment, or what physicists call taste.

The authors also see consequences for training and access. Problems that once suited early-career researchers may no longer be appropriate unless the goal is only to work through material. They expect costs to rise sharply for harder problems and say theoretical physics should discuss access and equal opportunity before capable agents reshape the field.

Sources: The Decoder