← All news·2026-10-16·3 min read

GPT-6-Astra Solves 49 of 50 Physics Research Problems

In the new TheoPhysEval benchmark, GPT-6-Astra solved 49 of 50 unpublished theoretical physics problems. Opus 5.5 came in second with 92 percent.

aiopenaisciencebenchmark

35 researchers, including scientists from Harvard, Belgium's KU Leuven, and Google's AI lab DeepMind, have collaborated on creating a benchmark of 50 unpublished theoretical physics questions called TheoPhysEval. OpenAI’s agent GPT-6-Astra has managed to solve 49 out of 50 questions (98%), spending less than $3 on average for each question and a total of just under $116 for all of them. In second place is Anthropic’s Opus 5.5 model which runs within their code writing assistant Claude Code, achieving a score of 92%. Trailing behind is Google’s Gemini 3.8 Flash, which continues to catch up, scoring 68%.

When restricted to not using the internet, both Astra and Opus achieve a score of 92%. According to the authors, this suggests that these models have essentially memorised the body of physics knowledge available, instead of thinking through the problem themselves. These are simple tasks that could be completed by an expert within a day, but they aren’t groundbreaking discoveries. What’s really interesting here is how cheaply we can get this kind of work done — it only takes a few dollars to replace an entire day of work for a physicist! But remember, you shouldn’t base your decision on which model to use solely based on a leaderboard result.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids