Artificial Intelligence Has Made a Historic Leap: New Models Solve Problems Beyond Even the Best Mathematicians

The latest advances in artificial intelligence are astonishing in their speed, exceeding even the boldest expectations. Volodymyr Bandura, CEO of Innolyticsgroup and ambassador of Singularity University, shares exclusive details about the new stage in the development of AI models. This is not just about improvements in solving mathematical tasks, but about a radical paradigm shift: the models are demonstrating the ability to independently find solutions to entirely new, previously unknown challenges, successfully passing tests specifically designed to prevent simple brute-force searching.

One of the developers of such advanced systems noted that future generations of models will be an “ontological shock” for users. These words already seem to be coming true. We are rapidly approaching the point where the range of tasks for which a human can clearly formulate both the question and the answer will begin to run out for AI systems.

Mathematical Puzzles: From Child’s Problems to Frontier Research

Just a few years ago, the difficulty of elementary school mathematics left most language models stumped. Experts often emphasized the limitations of such systems, calling them “stochastic parrots” — entities that only statistically imitate language but do not possess true logical reasoning. However, the emergence of models capable of “deep” thinking has radically changed the situation.

By 2025, at a prestigious international mathematics olympiad, artificial intelligence had won gold medals, competing on equal footing with humans under standard rules. This came as a surprise to many, but leading AI researchers had predicted such a development.

That is why, in late 2024 and early 2025, the Frontier Math benchmark was introduced — one of the most difficult mathematics test suites. Creating this test required dozens of leading mathematicians competing for the right to formulate hundreds of challenging problems. These tasks were divided into four difficulty levels: from university-level mathematics to advanced research problems. The fourth category is so difficult that top specialists could spend weeks developing a single problem.

To assess the benchmark’s difficulty and compare it with human capabilities, the authors tested it with the most talented students at the Massachusetts Institute of Technology (MIT). Eight teams of four to five people each had 4.5 hours to complete the test. The results were striking: on the relatively simple tasks in the first three levels, the teams scored an average of 19%, and all teams combined — 35%. The fourth-level tasks, even more complex, were not included in the student tests at all, since even leading mathematicians with Fields Medals admitted they could solve only a few of them, mostly within their narrow specialties.

At the beginning, the most advanced AI models at the time achieved less than 10% even on the easiest tasks in the first three Frontier Math levels. However, the new model, GPT-6 Astra, showed absolute dominance, solving 97.6% of the hardest, fourth-level tasks. This means the model handled virtually all problems that required genuine mathematical genius, demonstrating a level that far exceeds the capabilities of any living mathematician.

Practical Applications and New Horizons

This naturally raises the question: how do these impressive academic achievements translate into practical use? The answer lies in fields where complex mathematical models already play a key role. The clearest example is the financial conglomerate BlackRock, whose Aladdin software system manages assets worth more than $21 trillion. Originally developed for BlackRock, this system is now used by numerous other organizations.

In the context of rapid AI development, models that demonstrate a higher mathematical level than even the most experienced teams at giants like BlackRock open unprecedented opportunities. The potential to develop more advanced algorithms, analytical tools, and forecasting models is enormous. While this may spark debates about the influence of “secret world governments” or other conspiracy theories, facts remain facts: new AI models are creating the foundation for a qualitatively new level of development.

A Leap in Creativity and Symbolic Reasoning

The astonishing abilities of new models are confirmed not only in mathematics. Their progress in creative problem-solving tests, such as ARC AGI 3, is equally impressive. This benchmark is known for its difficulty and for its design, which punishes models for attempting brute-force search.

The previous best OpenAI model, GPT 5.6 Sol, scored only 7.8% on ARC AGI 3. The new model, Astra, achieved 99.9%, showing its ability to instantly grasp the logic of new tasks even when the rules are unknown. At the same time, Astra demonstrated more efficient solution paths than the average human, using fewer steps.

ARC AGI 3 researchers were extremely surprised by the level of symbolic reasoning demonstrated by Astra. The model created its own algebraic language for describing game states, which greatly simplified its problem-solving process. ARC AGI 3, introduced only at the end of March 2026, was so difficult at launch that no model exceeded a 10% score. However, that record lasted only a few months, unlike the first version of ARC, which remained unbeaten for five years.

These achievements are only the beginning. Dozens of similar benchmarks confirm the record-breaking performance of the new models. In just a few days, Astra will become available to the general public, which means we are standing on the threshold of a new era of artificial intelligence, where machine capabilities surpass the boldest human assumptions.