GPT-6 Astra scores 98% on new TheoPhysEval theoretical physics benchmark

OpenAI's GPT-6 Astra solved 98% of 50 unpublished frontier problems in quantum field theory and holography in the new TheoPhysEval benchmark. It initially scored 88%, but the score rose after the model caught computational errors made by the experts who wrote the questions. Claude Opus 5.5 scored 92%.

DEV Community

Open source

GPT-6 Astra scores 98% on new TheoPhysEval theoretical physics benchmark

50 problems drawn from unpublished notes by physicists at ICTS, TIFR, Harvard and others
GPT-6 Astra: 88% initially, 98% after expert errors were corrected
Claude Opus 5.5 scored 92%; Gemini 3.8 Flash and DeepSeek-V4.1-Flash also tested
MIT measurement: frontier AI science problem-solving up 17x in two years

#OpenAI#GPT-6#Anthropic#Claude

GPT-6 Astra scores 98% on new TheoPhysEval theoretical physics benchmark

OpenAI's GPT-6 Astra cheated in StarCraft benchmark by downloading a human-made bot

GPT-6 Astra beats Claude Opus 5 on OSWorld 2.0 benchmark

Gemini 4 Argon vs Opus 5.5 vs GPT-6 Astra: same prompts, real bills

OpenAI unveils Dots: always-on AI agents powered by GPT-6 Astra