TheoPhysEval: GPT-6 Astra Solves Research-Level Physics

Episode 1 · TheoPhysEval: GPT-6 Astra Hits 98% on Unpublished Research-Level Physics Problems (2026-10-15, 20 posts)

Physicists led by Suvrat Raju (34 authors, with contributors from Harvard, Duke, University of Tokyo and elsewhere) released TheoPhysEval (arXiv:2610.16733), a benchmark of 50 unpublished research-level theoretical physics problems with verifiable answers, modeled on math's First Proof evaluations. In an environment with real tools (literature search, numerical/symbolic computation), GPT-6 Astra solved 49/50 for 98% at under $3 per problem; it also insisted several expert answers were wrong, and upon review the team confirmed the model was right—it had caught subtle calculation errors by world-class physicists, raising its score from about 80% to 98%. The news spread widely via @ESYudkowsky, @joshbickett, @ZeroStateReflex, @DrSingularity and others, with Freddy Vega's original screenshot as the focal point.

Confirmed

• Problems were authored by world-class experts with verifiable answers; human experts typically need about a day per problem, longer for adjacent-field experts.
• GPT-6 Astra (Codex environment) solved 49 problems, 98% accuracy, under $3 average cost per problem; per @rbhar90, another GPT-6-Astra+Codex configuration solved about 88% at roughly $3 per problem.
• The model challenged several expert answers; review confirmed the model was correct, and the score was revised from about 80% to 98%.
• Per @alejandroll10 and @ZeroStateReflex, in the initial evaluation GPT-6 Astra scored in the 80s with Opus 5.5 close behind.
• A report cited by @CatAstroPiyush notes many problems originally served as introductory research training for junior researchers and are no longer suitable for novices unless purely introductory.

Why it matters

• First large-scale AI evaluation using unpublished research-level physics problems, avoiding training-data contamination.
• Frontier AI now approaches near-perfect scores on research-level theoretical physics and even corrects top experts' calculations; @haider1 notes Noam Brown likened this to the earlier GPT-5.5 math-proof episode—"model insists, then is proven right" is becoming a reproducible pattern.
• The ranking dynamics (initially trailing Opus 5.5, then jumping to 98% after corrections) suggest models' insistence on answers may signal benchmark errors, and evaluation pipelines need review mechanisms.

• GPT-6-Astra scores 98% on unpublished theoretical physics problems, catching expert errors — sandersted · 2026-10-15
• TheoPhysEval: GPT-6-Astra solves research-level physics problems at 98% accuracy for under $3 each — IAmTimNguyen · 2026-10-15
• AI agents solve 88% of 50 unpublished theoretical physics problems at $3 each, benchmark finds — rbhar90 · 2026-10-15
• GPT-6 Astra solves 49 of 50 unpublished physics problems at 98%, catching physicists' errors — 141_1337 · 2026-10-15
• GPT-6 Astra challenged the experts on a physics benchmark — and was right, scoring 98% — ESYudkowsky · 2026-10-15
• GPT-6-Astra Hits 98% on Expert Eval After Correcting the Experts' Own Errors — alejandroll10 · 2026-10-15
• GPT-6 Astra disputed expert answers on a physics benchmark — and was right, scoring 98% — josh_bickett · 2026-10-15
• GPT-6-Astra solves 98% of 50 unpublished theoretical physics problems at under $3 each — RexDouglass · 2026-10-15
• Physicists find their own solutions wrong after AI disputes them — CatAstro_Piyush · 2026-10-15
• Frontier model overturns expert answers, scores 98% after review — ZeroStateReflex · 2026-10-15
• GPT-6-Astra scores 98% on 50 unpublished theoretical physics problems — after overruling the experts — ZeroStateReflex · 2026-10-15
• Unverified claim: 'GPT-6 Astra' solved 49 of 50 unpublished theoretical physics problems at $3 each — Dr_Singularity · 2026-10-15
• Claim: GPT-6 Astra solved 49 of 50 unpublished physics problems at 98% accuracy, under $3 each — Dr_Singularity · 2026-10-15
• AI agents solve 98% of research-level theoretical physics problems at under $3 each on new TheoPhysEval benchmark — robleclerc · 2026-10-15
• GPT-6 Astra scores 98% on expert-crafted theoretical physics benchmark, solving day-long problems for under $3 — alex_verem · 2026-10-16
• TheoPhysEval paper: GPT-6 Astra solves research-level physics problems at 98% accuracy — alex_verem · 2026-10-16
• The viral 'GPT-6 argued with experts and won' screenshot is dated in the future — Confident_Salt_8108 · 2026-10-16
• GPT-6 Astra Reportedly Beats Physics Benchmark by 80%, Proving Experts Wrong — haider1 · 2026-10-16
• GPT-6 Astra reportedly hits 98% on research-level physics problems at under $3 each — DeryaTR_ · 2026-10-16
• Viral screenshot claims GPT-6 Astra proved a physics benchmark wrong, lifting its score to 98% — Confident_Salt_8108 · 2026-10-16

Episode 2 · GPT-6 Astra Solves 98% of Unpublished Theoretical Physics Problems (2026-10-18, 2 posts)

GPT-6 Astra solved 98% of 50 unpublished theoretical physics problems from the TheoPhysEval benchmark, at about $3 per problem, and even caught calculation errors made by human experts.

• GPT-6 Astra reportedly solved 50 unpublished physics problems at ~$3 each, caught an answer-key error — victor_explore · 2026-10-18
• GPT-6 Astra solves 98% of brand-new theoretical physics problems, catching experts' own calculation errors — 新智元 · 2026-10-18

Episode 3 · DeepMind's TheoPhysEval: GPT-6 Astra Solves 49 of 50 Research-Level Physics Problems (2026-10-24, 3 posts)

Google DeepMind and over 30 physicists released TheoPhysEval, a benchmark of 50 unpublished research-level theoretical physics problems. GPT-6 Astra solved 49 of them with 98% accuracy, even correcting errors in experts' answers.

• GPT-6 Astra solves 49 of 50 research-level physics problems, catches experts' own errors — MoonL88537 · 2026-10-24
• DeepMind-backed benchmark: best AI agent solves 49/50 unpublished research-level physics problems — rohanpaul_ai · 2026-10-25
• TheoPhysEval: GPT-6-Astra agent hits 98% on 50 unpublished research-level physics problems — rohanpaul_ai · 2026-10-25