OpenAI O3 Scores 96.7% on Math Olympiad Problems That Stump Most Humans

OpenAI's O3 model has achieved something remarkable in mathematical reasoning. On the American Invitational Mathematics Examination (AIME), one of the most challenging high school math competitions in the United States, O3 scored 96.7% accuracy on the 2024 version. That translates to solving 14.5 out of 15 problems correctly.

To understand how impressive this is, consider what AIME actually tests. These aren't standard textbook problems. The median human score is just 4 to 6 correct answers out of 15. Only the top 5% of students who take the AMC 12 (itself a challenging competition) even qualify to attempt AIME. And among those elite students, scoring above 10 out of 15 puts you in contention for the USA Mathematical Olympiad.

O3's 96.7% performance places it solidly in the top 1% of all human AIME competitors. But here's what makes the achievement more nuanced than it first appears: O3 isn't superhuman at math. It's human-level at the elite end, with specific strengths and surprising weaknesses.

Before diving into O3's performance, it's important to understand what AIME measures and why it matters as an AI benchmark.

AIME consists of 15 problems testing advanced high school mathematics across multiple domains:

Problem Types:

Key Characteristics:

The format is deliberately unforgiving. You can't guess your way to success. Each problem requires understanding the underlying mathematics, devising a solution strategy, executing multiple steps correctly, and arriving at a precise numerical answer.

The median score of 4 to 6 out of 15 isn't because students are poorly prepared. AIME qualifiers have already demonstrated exceptional mathematical ability by scoring in the top 5% on the AMC 12 or top 2.5% on the AMC 10.

What makes AIME challenging:

Top performers at AIME typically score 10 to 12 out of 15. Scoring 14 or 15 is exceptionally rare and usually indicates someone who will advance deep into International Mathematical Olympiad territory.

O3's 96.7% score on AIME 2024 represents 14.5 out of 15 problems solved correctly. But the story becomes more interesting when you examine performance across different versions and compare it to other AI models.

Independent testing revealed an important pattern. O3 performed significantly better on AIME 2024 (released in February 2024) than on AIME 2025 (released in February 2025).

Performance breakdown:

This raises questions about whether O3's training data included AIME 2024 problems or similar content. The performance gap suggests that when faced with truly novel problems, the model's accuracy drops.

According to independent benchmark testing from Vals.ai, no current model has fully mastered AIME 2025. The correctly answered questions are distributed across different models, suggesting that no single AI has developed a comprehensive approach to these problems consistently.

O3's performance represents a dramatic leap from earlier AI capabilities, but it's not the only model achieving high scores.

TABLE 1: AI Model AIME Performance Progression

The progression is striking. In less than a year, AI performance on AIME jumped from barely passing (42%) to near-perfect (96.7%). This represents one of the fastest capability gains on any mathematical reasoning benchmark.

O3 performs exceptionally well on certain types of AIME problems:

Strong Performance Areas:

These problem types play to O3's strengths: following multi-step logical procedures, maintaining consistency across calculations, and exploring solution spaces systematically.

Despite the impressive overall score, O3 shows weaknesses on specific problem types:

Weaker Performance Areas:

The difference matters. Top human mathematicians often solve AIME problems through elegant insights that short-circuit lengthy calculations. O3 tends toward brute-force systematic approaches, which work most of the time but fail on problems designed to reward creativity.

Placing O3's 96.7% score in human context reveals both its capabilities and limitations.

AIME results follow a predictable pattern among qualified students:

TABLE 2: AIME Human Performance Percentiles

O3's score of 14.5 places it firmly in the 98-99th percentile. It outperforms 99% of human AIME competitors, including many students who go on to study mathematics at top universities.

The key insight: O3 reaches elite human performance but doesn't exceed the very best humans.

Human advantages:

O3 has reached "really good human" level, not "superhuman" level. This distinction matters for understanding AI's current capabilities in mathematical reasoning.

The 96.7% score tells us something important about the state of AI reasoning models.

O3's success on AIME validates the reasoning model approach. Unlike earlier models that generated answers in one pass, O3 uses extended "thinking time" to work through problems step by step.

The process:

This mirrors how human mathematicians approach AIME problems, and the results show it works. The model can maintain logical consistency across complex multi-step solutions, a capability that eluded earlier AI systems.

Despite the reasoning capabilities, O3 likely relies heavily on pattern matching from training data.

Evidence for this:

The model has learned to recognize mathematical problem types and apply appropriate solution templates. This is useful but different from genuine mathematical understanding.

O3's exceptional performance raises questions about what mathematics appeared in its training data.

Potential training sources:

OpenAI hasn't disclosed exactly what mathematical content trained O3. The performance gap between AIME 2024 and 2025 suggests the model may have indirect exposure to problems similar to the 2024 exam through publicly available training data.

AIME performance, while impressive, reveals clear limitations in AI mathematical capabilities.

O3's 96.7% on AIME contrasts sharply with performance on more difficult mathematical challenges:

TABLE 3: O3 Performance Across Math Benchmarks

*December preview version; production version ~10%

The pattern is clear: as problems require more creativity, longer reasoning chains, and genuine mathematical insight, O3's performance degrades significantly.

The performance across benchmarks reveals important truths:

O3 can:

O3 cannot:

This mirrors findings across other domains. AI excels at systematic application of learned patterns but struggles with genuine creative reasoning.

O3's AIME performance matters beyond benchmark competitions.

AI systems reaching this level of mathematical competence create both opportunities and challenges for education:

Opportunities:

Challenges:

The fact that O3 can solve 96.7% of AIME problems means it can handle essentially all standard high school and undergraduate mathematics. This forces educators to rethink what and how they teach.

Mathematical problem-solving is central to many technical careers:

Impact on:

AI systems with AIME-level mathematical competence can assist with routine mathematical work in these fields. The key word is "assist." The creative mathematical thinking that drives breakthroughs remains human territory.

O3's performance provides crucial data about AI capability development:

What we learned:

This informs development priorities for next-generation AI systems. Improving reasoning capability shows clear returns. The remaining challenge is developing genuine mathematical creativity rather than pattern application.

AIME is approaching saturation as an AI benchmark. O3-mini already scores 86.5%, and newer models will likely push even higher.

The AI research community is already moving toward harder evaluations:

Emerging benchmarks:

These benchmarks test capabilities beyond AIME's structured problem format. They require not just calculation and logic but genuine mathematical creativity.

When AI systems consistently score 100% on AIME, it will represent an important milestone:

But it won't mean:

AIME tests a specific, valuable form of mathematical reasoning. Perfect performance on it demonstrates capability at that level, nothing more.

OpenAI's O3 scoring 96.7% on AIME represents genuine progress in AI mathematical reasoning. The model has reached elite human performance on problems that stump 99% of students, demonstrating that multi-step logical reasoning is teachable to AI systems.

But context matters. O3 performs better on older AIME exams than newer ones, suggesting training data exposure. It excels at systematic problem-solving but struggles with creative insights. And its performance drops dramatically on harder benchmarks requiring genuine mathematical creativity.

The achievement is impressive within its proper scope. O3 can reliably solve complex, multi-step mathematical problems that require maintaining logical consistency across extended reasoning chains. This has clear applications for education, technical work, and advancing AI capabilities.

What it doesn't demonstrate is mathematical understanding equivalent to top human mathematicians. The best humans still solve AIME problems faster, more creatively, and with perfect accuracy. And on research-level mathematics, AI systems still lag far behind expert human performance.

O3's 96.7% on AIME marks an important milestone in AI development. It shows that AI has mastered a significant portion of mathematical problem-solving. But the gap between mastering structured olympiad problems and achieving genuine mathematical creativity remains vast.

For now, O3 represents an extremely capable mathematical assistant rather than a mathematical innovator. That's valuable enough to transform how we approach mathematical education and technical work. But it's not the end of human mathematical reasoning. It's the beginning of a new phase where humans and AI collaborate on mathematical problems, each contributing their distinct strengths.