From its earliest ambitions, artificial intelligence has carried a bold promise: not just to automate tasks, but to expand the boundaries of human knowledge itself. Demis Hassabis founded DeepMind with the goal of “solving intelligence” and then “solving everything else.” Sam Altman has repeatedly argued that AI-driven scientific acceleration could deliver “enormous” gains to global quality of life. Anthropic CEO Dario Amodei has gone even further, predicting that by 2026 AI could amount to a “country of geniuses in a data center.”
This belief—that AI could fundamentally transform how humanity understands the universe—has become one of the most enduring myths, and motivations, behind the current AI boom. A new benchmark from OpenAI, called FrontierScience, offers fresh evidence that this vision may not be entirely speculative.
Measuring AI at the Frontiers of Science
Published this week, FrontierScience is designed to rigorously evaluate whether advanced AI models can meaningfully contribute to scientific research. According to Miles Wang, a researcher on OpenAI’s evaluation team who led the project, the goal is clear:
“We want to rigorously measure how models can improve scientific capabilities and maybe even accelerate scientific discovery.”
The benchmark spans physics, chemistry, and biology, and is divided into two levels:
- Olympiad Tier:
Contains 100 highly challenging questions comparable to international science Olympiads, designed to test the limits of elite human problem-solving. - Research Tier:
Comprises 60 open-ended questions written by PhD-level scientists, aimed at evaluating reasoning, judgment, and real-world research support.
These research-tier problems are not trivial. One question on meso-nitrogen atoms in nickel(II) phthalocyanine would require multi-day computer simulations to solve accurately. Another, involving electrostatic wave modes in plasma, reflects analyses that can take experienced researchers weeks to complete.
Rapid Progress, Rising Scores
The results from FrontierScience show a familiar pattern across modern AI research: rapid improvement.
OpenAI’s newly released GPT-5.2 currently leads the benchmark, scoring:
- 77.1% on the Olympiad Tier
- 25.3% on the Research Tier
While the research-tier improvement over GPT-5 is modest, Wang notes that the pace of advancement has been striking.
“Progress has been intensely fast over the last year with reinforcement learning and reasoning models.”
To put this into perspective, AI benchmark performance in many reasoning-heavy domains has increased by 30–50 percentage points in just two years, according to aggregated evaluations from AI research organizations. FrontierScience appears to follow this same upward trajectory.
If models eventually approach near-perfect performance on research-tier questions, Wang believes they could become powerful collaborators—multiplying the productivity of PhD students and professional scientists rather than replacing them.
The Limits of Benchmarking Scientific Intelligence
Despite its promise, FrontierScience has important limitations. The benchmark is text-only, meaning models are not evaluated on experimental design, lab execution, or image and video analysis—all critical components of real scientific work.
Additionally, the relatively small question sets make it difficult to draw fine-grained comparisons between closely matched models. Notably, there is also no published human baseline to show how expert scientists would score under the same conditions.
Jaime Sevilla, director of research at Epoch AI, describes this as an unavoidable challenge:
“It’s very hard to design a benchmark that’s truly informative about when models will be useful in real research.”
As AI systems approach expert-level reasoning, even evaluating them becomes increasingly expensive and difficult. When benchmark questions are written by world-leading experts, finding independent evaluators with equivalent expertise is often impractical.
The Growing Industry Behind AI Evaluation
This challenge has fueled the rise of specialized expert data companies such as Surge AI and Mercor, both reportedly valued at over $10 billion. These firms recruit domain experts from academia and industry to design questions and grading rubrics for advanced AI systems.
Edwin Chen, founder and CEO of Surge AI, frames the mission bluntly:
“If you want to see the Riemann hypothesis proved in your lifetime, you’re going to help train an AI to either solve it or collaborate on solving it.”
AI’s Real Impact on Science—So Far
AI has already delivered tangible scientific breakthroughs, though mostly in narrow domains:
- AlphaFold, developed by Google DeepMind, has predicted over 200 million protein structures, a task that would have taken hundreds of millions of years using experimental methods alone.
- AI systems are now used to simulate plasma behavior in nuclear fusion reactors.
- Advanced models generate highly detailed, short-term weather forecasts with accuracy improvements of 20–30% over traditional systems.
Yet these tools typically address isolated stages of the scientific process. As researchers note, AlphaFold predicts protein structures—but not electronic properties, functional behavior, or experimental outcomes.
The broader ambition is an AI that can assist across the entire scientific lifecycle: forming hypotheses, designing experiments, analyzing results, and synthesizing insights across disciplines.
Promise, Productivity, and Pushback
Large language models are beginning to demonstrate that level of generality in mathematics and software development. OpenAI researcher and mathematician Sébastien Bubeck recounts giving GPT-5 a mathematical problem that had stumped his team for years. After two days of reasoning, the model discovered a key identity and solved it.
In applied research, productivity gains are already visible. Coding tasks that once took four hours now take some scientists 30 minutes or less. According to surveys conducted in 2024, researchers using AI-assisted coding report productivity increases of 25–40%.
Still, skepticism remains strong. Theoretical physicist Carlo Rovelli warns that large language models frequently generate confident but incorrect outputs. He also notes that AI-generated manuscripts have contributed to a doubling of journal submissions in some fields—many of them low-quality.
A Future of Collaboration, Not Replacement
Despite the criticism, few researchers dismiss AI outright. Instead, many view the technology as an evolving assistant rather than an autonomous discoverer.
If the trend identified by FrontierScience continues, AI systems may soon become reliable collaborators—handling calculations, simulations, and literature synthesis at unprecedented scale. That prospect leaves scientists both excited and overwhelmed.
As one chemist put it, reflecting on the pace of change:
“Too many feelings. I need a LLM to summarize them.”
For now, FrontierScience offers a glimpse into a future where AI doesn’t just answer questions—but helps humanity ask better ones.
Read more: OpenAI CCO Hannah Wong Announces Departure After Five Years
