Mathematics is the domain where reasoning models most clearly outperform standard language models. On competition-level benchmarks, the performance gap between reasoning models and standard models is larger than in almost any other domain — reasoning models score 70–95% on problems that defeat standard models at 20–40%. This is not a coincidence: mathematics is precisely the kind of task where intermediate reasoning steps are checkable, where errors in one step compound into wrong final answers, and where extended self-verification consistently improves accuracy. Understanding the math benchmark landscape and what reasoning models can and cannot do in mathematics shapes realistic expectations for applying them to quantitative work.
The Key Math Benchmarks
Three benchmarks define the performance landscape for mathematical reasoning models. MATH (Hendrycks et al.) is a dataset of 12,500 competition mathematics problems across seven difficulty levels, from introductory to competition-level. Standard models score around 50–60% on MATH-500 (a representative 500-problem subset); reasoning models score 95–97%. AIME (American Invitational Mathematics Examination) is the benchmark that matters most for distinguishing frontier models — these are genuinely hard competition problems that require real mathematical insight. A score of 15/15 would represent solving every problem in an AIME paper; strong human competitors average around 5–8. o3 scored approximately 25.2/30 (across both AIME 2024 papers), which would place it at the level of a very strong competition finalist. DeepSeek R1 scored approximately 79% on AIME 2024. GPQA Diamond (Graduate-Level Google-Proof Q&A) covers PhD-level science, including physics and chemistry problems that require multi-step mathematical reasoning — reasoning models score around 75–85%, compared to 50–60% for standard models.
Why Mathematics Is the Natural Home for Reasoning Models
Mathematics is ideally suited to the reasoning model paradigm for structural reasons. Every step in a proof or calculation can be verified: if you write “x = 5” in step 3 and your reasoning depends on “x = 7” in step 5, the error is detectable. The model can substitute its answer back into the original equation to check it. It can try to prove the same statement two different ways and verify they agree. It can check whether a derived formula produces the right answer on small numerical examples before generalising. None of these verification strategies require external information — they are self-contained logical checks that the model can perform using only its thinking tokens. This is why extended thinking helps mathematics more than almost any other domain: the model has internal checkpoints it can use to catch and correct errors before finalising an answer.
AIME Performance in Detail
AIME problems deserve special attention because they are the benchmark that most clearly separates frontier reasoning models from earlier generations. These problems require sustained multi-step reasoning with no partial credit — you either identify the key insight and execute the solution correctly, or you do not. Common problem types include combinatorics (counting arrangements with specific properties), number theory (problems involving divisibility, modular arithmetic, and properties of integers), geometry (often requiring careful algebraic manipulation of geometric relationships), and algebra (polynomial identities, function equations, optimization). o3 solving the majority of these problems correctly places it in a genuinely useful tier for competition preparation, mathematical research assistance, and checking proofs in applied mathematics — tasks that were simply unavailable from AI systems two years earlier.
Figure 1 — Math benchmark scores: reasoning models vs standard models
What “95% on MATH-500” Actually Means
MATH-500 scores above 95% sound remarkable but need context. MATH problems, even at the hardest levels (Level 5), are well-represented in training data — many of these exact problems, or problems of the same type, have appeared in textbooks, forums, and online resources that the models were trained on. High MATH-500 scores reflect both genuine mathematical reasoning capability and memorisation of problem types and solutions. AIME is a better discriminator because AIME problems are released annually and are specifically designed to be novel — they reward mathematical insight that cannot be retrieved from memory. The jump from 9% (GPT-4o on AIME) to 84% (o3 on AIME) is therefore the more meaningful number: it represents genuine reasoning capability improvement, not just better memorisation.
Practical Applications for Quantitative Work
For practitioners doing quantitative work, reasoning models open up several practically useful capabilities. Deriving formulas: given a physical or statistical setup, reasoning models can derive the correct formula by working through first principles — a task that standard models frequently get wrong by confusing similar-looking formulas. Checking proofs: ask a reasoning model to verify whether a given mathematical argument is correct, and it will systematically trace through each step. This is useful for catching errors in analytical work before they propagate. Solving optimisation problems: constrained optimisation problems that require Lagrange multipliers, KKT conditions, or careful case analysis are handled well by reasoning models that can work through the conditions systematically. Statistical reasoning: problems involving conditional probability, Bayesian updating, or hypothesis testing under specific conditions are handled much more reliably by reasoning models than standard ones — these are exactly the kind of multi-step logical tasks where standard models make systematic errors.
Where Reasoning Models Still Struggle in Math
Despite impressive benchmark performance, there are systematic limitations in current reasoning models’ mathematical capabilities. Long formal proofs: proving a non-trivial theorem from scratch — not solving a competition problem but writing a rigorous proof with all steps justified — remains unreliable. Models can produce proofs that look correct but contain subtle logical gaps. Novel research mathematics: problems at the frontier of mathematical research, where no similar problem exists in training data, expose the limits of pattern-based reasoning even with extended thinking. Numerical computation: reasoning models are not calculators and can still make arithmetic errors, especially in long computation chains. For important quantitative results, verify numerical computations independently. Symbolic manipulation: very long algebraic manipulations with many terms tend to accumulate errors — for symbolic computation, a computer algebra system (Mathematica, SymPy) is more reliable than a language model regardless of its reasoning capability.
Reasoning Models as Math Tutors
One of the most practically valuable applications of reasoning models’ mathematical capability is tutoring and explanation. Because reasoning models can work through problems step by step, they can explain not just the answer but the reasoning process in a way that builds mathematical intuition. When a reasoning model solves an AIME problem, reading the thinking trace shows the mathematical moves it makes — noticing a symmetry, choosing a clever substitution, recognising that a problem reduces to a known form. For students learning competition mathematics, this provides the kind of detailed problem-solving process that previously only came from working through solutions with an experienced mentor. The quality of these explanations is substantially better when the model reasons explicitly rather than pattern-matching to a memorised solution.
The Frontier: IMO-Level Problems
The International Mathematical Olympiad (IMO) represents the hardest competition mathematics — problems that take the world’s best student mathematicians hours to solve, if they can solve them at all. As of mid-2026, reasoning models can solve some IMO problems but not reliably — they might solve 1–2 out of 6 problems on a given paper, roughly comparable to a strong national team qualifier but far below a gold medallist. The problems that stump them most consistently are those requiring truly novel insight — noticing a non-obvious connection between two mathematical objects, or constructing an auxiliary object whose existence is not suggested by the problem statement. These are the hardest class of mathematical reasoning tasks, and they represent the current ceiling of what test-time compute scaling can achieve. Progress is rapid, however: what was IMO-impossible for AI systems in 2023 is now occasionally solvable by frontier reasoning models in 2026.
Reasoning models represent a genuine breakthrough in mathematical capability. The jump from 9% to 84% on AIME between standard and frontier reasoning models is not incremental — it is a qualitative change in what is possible. For applied quantitative work — formula derivation, proof checking, statistical reasoning, optimisation problem setup — reasoning models are now genuinely useful tools rather than unreliable assistants. The limitations are real but bounded: arithmetic errors in long computations, gaps in formal proofs, and failure on truly novel research problems. Knowing where the boundaries are allows you to use these capabilities where they are reliable and verify the outputs where they are not.