GPT-6-Astra Solves 49 of 50 Unpublished Physics Problems in Landmark Benchmark

An AI model just worked through a stack of 50 unpublished, research-level theoretical physics problems. It solved 49 of them. That result comes from a new preprint called TheoPhysEval, and the model in question is GPT-6-Astra. With web search turned on, it scored 98%, the kind of grade that usually gets a graduate student a handshake and a heavier workload.

The paper titled "TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving" landed on arXiv on October 9, 2026. The benchmark contains 50 theoretical physics problems that had never been published, spanning 11 subfields including fluid dynamics and quantum field theory. The goal was to see what AI agents can do with no human help at all—nobody nudged the models along or fixed their mistakes midstream.

Physicists from institutions such as ICTS and Google DeepMind led the work. The author list runs to 35 people from a range of research institutions. The researchers tested off-the-shelf models, not custom-built research systems, and designed the criteria to approximate the skills an expert theoretical physicist would need. The study split testing into two tracks: the "realistic" track let models search the web, the way a working researcher would, while the "isolated" track cut them off from the internet entirely.

On the realistic track, GPT-6-Astra solved 49 of the 50 problems. The cost of that run came to $2,800, a figure that underscores the computational demands of such tasks. The isolated track saw lower performance, with models achieving roughly 60% accuracy when denied web access, highlighting the importance of real-time information retrieval in scientific problem-solving.

The benchmark represents a shift in how artificial intelligence is evaluated in scientific domains. Rather than testing models on textbook problems or rehashed exam questions, TheoPhysEval uses novel challenges that mirror the open-ended nature of theoretical physics research. This approach aligns with broader trends in AI development, where benchmarks are increasingly focused on real-world applicability rather than narrow metrics.

The market reaction to this development has been swift. As noted by TechCrunch, AI agents are rapidly expanding beyond text generation into specialized domains like scientific reasoning. The success of GPT-6-Astra on TheoPhysEval suggests that future iterations of AI models may become indispensable tools for researchers, particularly in fields requiring complex mathematical analysis and interdisciplinary knowledge.

The implications extend beyond physics. If AI can autonomously solve theoretical problems, it may accelerate progress in other domains such as artificial intelligence analysis in medicine or artificial intelligence as chemical trespass in materials science. However, as Anthropic’s recent article on "Claude-shaped science" argues, AI’s strength lies in computation and interdisciplinary work, while human expertise remains essential for determining the significance of results.

TheoPhysEval’s design reflects this division of labor. By creating problems that require reading literature, writing code, and verifying answers through multiple methods, the benchmark tests AI’s ability to function as a collaborative tool rather than a standalone researcher. This approach mirrors the philosophy behind projects like BootLoops, which organize scientific workflows to leverage AI’s strengths while maintaining human oversight.

The cost of running GPT-6-Astra on TheoPhysEval—$2,800—raises questions about scalability. While the model achieved near-perfect accuracy, the financial and computational resources required may limit widespread adoption. For now, such benchmarks serve as proof-of-concept demonstrations rather than practical tools for everyday research.

The artificial intelligence community is watching closely. As Evertune’s AI Model Tracker notes, the pace of model releases has accelerated dramatically, with 149 updates tracked as of October 21, 2026. Each new model brings incremental improvements, but TheoPhysEval represents a qualitative leap in AI’s capacity to engage with theoretical science.

What does this mean for researchers? For now, AI agents like GPT-6-Astra are best viewed as assistants rather than replacements. They can process vast amounts of information, run simulations, and even propose hypotheses—but the interpretation and validation of results still require human judgment. The line between AI as tool and AI as collaborator continues to blur, raising both opportunities and challenges for the future of scientific discovery.

FAQs:
What is TheoPhysEval? It is a benchmark of 50 unpublished theoretical physics problems designed to test AI agents’ ability to solve research-level challenges autonomously.
How did GPT-6-Astra perform? It solved 49 of 50 problems on the realistic track (with web search) and achieved roughly 60% accuracy on the isolated track (without web access).
What does this mean for AI in science? It demonstrates AI’s potential to assist researchers in complex problem-solving, though human expertise remains critical for interpreting results and determining significance.
Is this a sign of AGI? Not yet. While impressive, GPT-6-Astra’s performance is task-specific and does not indicate general intelligence.