- Who it helps
- Scientists and readers assessing claims about AI research capability.
- Availability
- The preprint was submitted October 9. This brief is based on its abstract and the later discussion that surfaced it; benchmark artifacts and full experimental procedures still need closer inspection.
The reported result
The authors report 98% accuracy for their strongest tested configuration at an average cost below $3 per problem. The system could use search and scripts. These are results under the paper’s setup, rather than a general claim that an AI can carry out an entire research project.
The remaining scientific question
The paper distinguishes solving specified calculations from finding and validating an original breakthrough. A small accuracy drop without search also leaves questions about prior exposure. Independent evaluation should preserve the exact tasks, tool permissions and scoring rules.
What you can do
- Read the full paper and artifact availability before comparing a model against this result. Keep the distinction between solving a defined problem and choosing a valuable new research question visible. Open reference ↗
Where the evidence stops
Only the abstract and the linked commentary were inspected. Problems, scoring, code and results were not independently audited or reproduced.
Sources read. Sources checked October 19, 2026 UTC. No external product workflow, model inference or third-party project was executed.
Sources & further reading
Paper · paper
TheoPhysEval paper abstract ↗
Primary explanation read on October 19. Product operation and performance were not independently tested.
TheoPhysEval authors · page checked 2026-10-19 · Published 2026-10-09
Web · analysis
Import AI 477: scientific-capability argument ↗
Read the opening argument through the theoretical-physics evaluation example. The author’s economic forecasts remain opinions; primary benchmark evidence was followed separately.
Jack Clark · page checked 2026-10-19
Coverage note: The cited primary material was read on October 19. A matched original X post, a watched video comparison and independent execution remain unavailable.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
TheoPhysEval evaluates AI systems on fifty unpublished theoretical-physics problems. It examines whether strong benchmark performance extends to calculations created outside familiar public test sets.