Observed facts: The arXiv preprint describes 50 unpublished research-level theoretical-physics problems with verifiable answers, assembled by 35 researchers. It reports that GPT-6-Astra harnessed in Codex achieved 98% final-answer accuracy with an average cost under $3 per problem, using tools that included literature search and numerical or symbolic scripting, without human hints.
Mechanism: The benchmark turns a claim about research capability into a score over problems with checkable answers. The authors also report only small average accuracy changes when web search and background papers were withheld, and suggest significant memorization of literature. Why now: This result gives teams a concrete performance-and-cost datapoint while renewed claims about AI-assisted research make evaluation design consequential.
Implications: A high score can support the narrower claim that an agent handles this set of calculations under the tested conditions. It does not show that the system selects important open questions, creates durable new theory, or produces breakthroughs. Evidence boundary: The available evidence is an arXiv preprint's abstract; it does not independently establish robustness beyond the 50-problem suite or the causes of the limited change under ablation.
Watch condition: Watch for independent replication on new, held-out problems and detailed analysis distinguishing memorized results from genuinely novel reasoning. The authors' paper and any follow-up should clarify how answers were verified and how representative the tasks are.
This is a distinct analytical record because no prior New Runtime record was supplied for this benchmark; no related record is assumed.