GPT-6-Astra Tackles Complex Physics Research Tasks at Under $3 Each

Such assignments are traditionally given to doctoral students embarking on their first research projects. The baseline complexity remains considerable. For instance, one problem involved adapting equations from a five-dimensional black hole framework into four-dimensional spacetime, then running a custom script to obtain the final values. Coordinator Suvrat Raju, who co-authored the original paper, required roughly a full day to complete it. The study authors estimated that prior to the emergence of modern autonomous agents, only a small group of specialized physicists possessed the combined domain expertise and computational background to carry it out, typically needing several days.

Performance gaps among the models became apparent during these multi-step derivations. When attempting the black hole problem with search tools enabled, both Gemini and GLM copied five-dimensional formulas directly from the source paper without making the necessary adjustments for four dimensions. Interestingly, web access proved largely non-essential for top performers: running without internet connectivity, both GPT-6-Astra and Opus 5.5 maintained a 92 percent resolution rate, indicating that frontier models have effectively absorbed substantial portions of the theoretical physics corpus. Furthermore, multiple initial failures were traced back to subtle oversights in the human authors' reference answers rather than model hallucination; once corrected, Astra's responses were verified as accurate.

Calculation prowess versus scientific intuition

Despite these high calculation scores, theoretical physics has not yet seen the sort of landmark AI breakthroughs observed in pure mathematics, where systems have contributed to major milestones like the Navier-Stokes problem mentioned by OpenAI. The study's authors argue that numerical or symbolic derivations represent merely the final phase of research. Before calculation begins, physicists must formulate theories, identify productive inquiries, and pivot toward simpler, equally informative problems when an initial route fails.

Current systems lack the subjective scientific "taste" that human researchers rely on to distinguish worthwhile directions from tedious detours. When prompted to construct novel questions, the agents tended to fixate on uninteresting technicalities. Consequently, the long-standing pragmatic ethos in theoretical disciplines to "shut up and calculate" may invert: as mechanics become automated, conceptual framing and discernment will matter far more.

This shift carries educational and economic implications. Standard entry-level assignments for early-stage PhD candidates are becoming redundant as practice exercises unless intended purely as pedagogical drills. At the same time, while baseline tasks remain inexpensive, more demanding calculations will require substantially higher computing budgets. This could introduce financial disparities to theoretical physics—a domain historically protected by its low overhead—echoing debates on access currently seen in mathematics around the Leiden Declaration.

What it means for developers

For engineering teams building autonomous research pipelines, the TheoPhysEval results demonstrate that agentic workflows combining language models with code-execution environments can reliably execute intricate, multi-step symbolic tasks. However, performance sharply diverges across model tiers, showing that complex formal logic and domain-specific translation demand top-of-the-line foundation models.

Developers evaluating model choices for analytical pipelines can access systems from OpenAI, Anthropic, Google, and DeepSeek through a unified pay-per-token interface at https://apixoai.online without managing multiple platform accounts. Because benchmark scores can be skewed by human answer-key errors—a factor that has artificially depressed scores on older datasets like CritPt—developers testing reasoning systems should incorporate independent verification and sandboxed computational execution into their evaluation harnesses.

Source: GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece — The Decoder. Written by the Apixo team from that report.