Skip to main content
BenchLM
Data

Best LLM for Math 2026: FrontierMath, AIME & HMMT Rankings

Best LLM for math 2026: live FrontierMath, AIME 2026 and HMMT February 2026 tables, why AIME 2025 and MATH-500 no longer pick a winner, and picks by use case.

Best LLM for Math 2026: FrontierMath, AIME & HMMT Rankings article cover
In this article7 sections

As of November 2026, Claude Fable 5.2 (95.088%, Benchmark exact) leads FrontierMath tiers 1-3, the hardest math test that most frontier labs still report. On fresh competition problems, GLM-5.2 (99.2%, Provider exact) leads AIME 2026 and Qwen3.7 Max (97.1%, Provider exact) leads HMMT February 2026.

Those are different questions, and the answer to "which LLM is best at math" depends on which one you are asking. The tables below rebuild from our math data on every site update, so the leaders named here are current. For the sortable view, use the math leaderboard.

Best LLM for math, ranked

FrontierMath tiers 1-3: the frontier separator

Table 1
Rank Model Score Evidence
1 Claude Fable 5.2 95.088% Benchmark exact
2 GPT-6 Astra 93.684% Benchmark exact
3 GPT-6.1 Sol 93.684% Benchmark exact
4 GPT-6.2 Astra 92.632% Benchmark exact
5 Claude Opus 5.5 91.228% Benchmark exact
6 GPT-6.1 Terra 90.877% Benchmark exact
7 Gemini 4 Carbon 90.526% Benchmark exact
8 Claude Fable 5.1 90.175% Benchmark exact
9 GPT-6 Sol 89.825% Benchmark exact
10 GPT-5.6 Sol 89.000% Provider exact

FrontierMath uses unpublished problems written by mathematicians, graded automatically, so a high score is hard to get from memorized training data. It is the table to read when you need the strongest model for hard algebra, number theory, combinatorics or analysis. The top rows are a few points apart; at that gap, price and latency decide more than rank.

FrontierMath Tier 4: research-level problems

Table 2
Rank Model Score Evidence
1 Claude Fable 5.2 100.000% Benchmark exact
2 GPT-6.1 Sol 100.000% Benchmark exact
3 GPT-6.2 Astra 100.000% Benchmark exact
4 GPT-6 Astra 97.600% Provider exact
5 Gemini 4 Carbon 95.122% Benchmark exact

Tier 4 is the research-grade slice. Fewer models report it, and its sample is small, so a single solved problem moves a score by several points.

AIME 2026 and HMMT February 2026: fresh competition math

Table 3
Rank Model Score Evidence
1 GLM-5.2 99.2% Provider exact
2 Beam 1.0 501B-A23B FP8 98.0% Provider exact
3 Beam 1.0 501B-A23B 97.7% Provider exact
4 Beam 1.0 501B-A23B NVFP4 97.5% Provider exact
5 A.X K2 97.1% Provider exact
6 Inkling 97.1% Provider exact
7 Kimi K2.6 96.4% Provider exact
8 Ternary Bonsai 2 27B 95.8% Provider exact
Table 4
Rank Model Score Evidence
1 Qwen3.7 Max 97.1% Provider exact
2 Inkling 2 Small 95.8% Provider exact
3 DeepSeek V4 Pro 95.2% Provider exact
4 DeepSeek V4 Flash 0731 94.8% Provider exact
5 DeepSeek V4 Pro (High) 94.0% Provider exact
6 Solar Open 2 93.9% Provider exact
7 Qwen3.7 Plus 92.9% Provider exact
8 Kimi K2.6 92.7% Provider exact

The 2026 competition sets were released after most training cutoffs, which makes them cleaner than the 2023-2025 papers. They are also close to saturated, and frontier labs largely stopped reporting them, so these two tables are led by open-weight and smaller labs. Read them as "which models clear competition math", not as a ranking of the frontier.

Why AIME 2025 and MATH-500 no longer pick a winner

When this post was first published on May 12, 2026, GPT-5.4 led it on AIME 2025 (99), MATH-500 (99) and BRUMO 2025 (97), with GPT-5.2 Pro tied on every cell and a dozen models within two points. That crowding is the problem. A one- or two-point gap on a benchmark where the top ten all score above 95 is noise until your own evaluation says otherwise, and the newer flagships do not report those rows at all: labs moved to FrontierMath, HMMT February 2026, IMOAnswerBench and APEX.

The older rows still have one use: contamination checks. A model that scores 99 on AIME 2023 but drops sharply on AIME 2025 or AIME 2026 probably memorized older public problems. A model that holds its score across years is showing math ability rather than recall. For the long version, read the AIME and HMMT explainer.

What each AI math benchmark tests

AIME

AIME, the American Invitational Mathematics Examination, is a 15-problem, 3-hour US high school competition with integer answers. It tests creative insight more than advanced notation, which is why it became a useful AIME LLM benchmark in the first place. We track AIME 2023, AIME 2024, and AIME 2025 to catch year-over-year drops that suggest memorization rather than general math ability.

HMMT

HMMT, the Harvard-MIT Mathematics Tournament, is comparable to AIME but more proof-flavored and multi-step. The strongest reported rows sit in the mid-to-high 90s, so it is still useful as a floor check but weak as a frontier separator. We keep HMMT 2023, HMMT 2024, and HMMT 2025 visible for the same contamination-check reason.

MATH-500

MATH-500 samples 500 problems from the Hendrycks MATH dataset across algebra, geometry, number theory, probability, and precalculus-style reasoning. Its broader difficulty range makes it more useful for mid-tier separation than AIME alone. DeepSeek-R1's 97.3 row is the clearest example: the model remains strong on MATH-500 even though its AIME 2025 and HMMT 2025 scores trail current frontier models badly.

BRUMO 2025

BRUMO 2025, the Bulgarian Mathematical Olympiad row, is one of the newer discriminators in our math set. The top models reach 96-97, but mid-tier reasoners drop faster than they do on AIME. Its 2025 date also lowers training-data contamination risk compared with older public problem sets.

FrontierMath, HMMT Feb 2026, IMOAnswerBench, and APEX

The next-gen suite exists because the older benchmarks are crowded at the top. FrontierMath uses research-grade problems written by mathematicians; the v2 split separates tiers 1-3 from the research-level Tier 4. HMMT Feb 2026 and AIME 2026 refresh the competition format with problems released after most training cutoffs. IMOAnswerBench uses Olympiad-style answer checking, and APEX exposes much steeper drop-off even for strong reasoners. For the general benchmark methodology, read what benchmarks measure.

Best LLM for math by use case

Competition math: AIME, HMMT, olympiad-style

Start from the AIME 2026 and HMMT February 2026 tables above; any model in their top rows clears this category. Require step-by-step reasoning or a reasoning mode, since even strong models drop sharply when forced to answer directly, and check the final answer against a second model when the stakes are high.

Engineering and applied math

Calculus, linear algebra, differential equations and numerical methods reward a model that writes and runs code. Pick from the top of the FrontierMath table, then confirm it on SciCode or your own scientific-computing tasks. A model that hands the arithmetic to a Python tool usually beats a slightly stronger model that does it in prose.

Finance and quantitative reasoning

For probability, expected value and model checking, numerical hallucination matters more than peak benchmark score. Prefer a model that states its assumptions and shows intermediate steps you can audit, and test it on your own spreadsheets or pricing models before trusting a one-shot answer.

Math tutoring and education

Tutoring is an explanation problem layered on top of a math problem. Every model in the top tables can solve school and undergraduate math; what separates them for teaching is whether the worked solution is clear, paced, and correct at every step. Run the same five problems through two or three candidates and pick the explanations students follow. No model is fully reliable, so verify worked solutions.

Math research and frontier reasoning

Claude Fable 5.2 (100.000%, Benchmark exact) leads the research-level FrontierMath Tier 4 table. The best models now solve most of its problems under benchmark conditions, which is real progress, but a benchmark answer is not a checked proof. Use these models to explore, suggest lemmas and find counterexamples, and verify anything you publish, ideally in a proof assistant. For related graduate-level coverage, compare HLE and the HLE explainer.

Math model pricing

Benchmark rank and price are separate decisions. The LLM pricing page lists current per-token rates for every model in the tables above, and the 2026 token-pricing guide explains how input, output and cached rates add up. For a few hard problems, use the strongest FrontierMath row you can access. For a high-volume math assistant, start lower in the table and move up only when your own error rate says so.

How we rank math models

The tables in this post are not a blended score. Each one sorts a single benchmark, and the Evidence column says where the number came from: "Benchmark exact" rows come from the benchmark publisher's own run, "Provider exact" rows from the model maker's report. A blended math score hides that difference, and on math it matters: the competition sets are mostly provider-reported, while FrontierMath is mostly run by its publisher.

Tracking three AIME years matters because it helps detect contamination. A model that scores 99 on AIME 2023 but collapses on AIME 2025 probably memorized older public problems. Frontier models score consistently across the three AIME rows, which is evidence of generalizable math reasoning rather than benchmark recall.

The full scoring model lives on the methodology page, and the sortable view of every math benchmark is the math leaderboard.

Which math table to trust for your task

Use FrontierMath tiers 1-3 to find the strongest model, Tier 4 for research-level problems, and AIME 2026 or HMMT February 2026 to confirm a model clears competition math. Use the live math leaderboard for sortable scores, the AIME and HMMT explainer for benchmark context, and the open-weight rankings if you need a self-hosted math model.

Frequently asked questions

01What is the best LLM for math in 2026?

Read the FrontierMath tiers 1-3 table at the top of this post: it is the hardest math test most frontier labs still report, and its first row is the strongest math model on our board. In October 2026 OpenAI's GPT-6 line and Anthropic's Claude Opus line held the top rows. Competition sets such as AIME 2026 and HMMT February 2026 are close to saturated, so treat them as a floor check.

02Which AI model is best at AIME?

The live AIME 2026 table in this post lists the top provider-reported rows. Most frontier labs stopped reporting AIME once scores crowded above 95, so the table is led by open-weight and smaller labs. At that level, pick on cost, latency and explanation quality rather than a one-point gap.

03Are AIME 2025 and MATH-500 still useful for choosing a math model?

Only as a floor check. Frontier models score near the ceiling on both, and several new flagships do not report them at all. Use FrontierMath for capability at the top end, AIME 2026 and HMMT February 2026 for fresh competition problems, and your own problem set for the final call.

04What is the best open-weight LLM for math?

Open-weight labs mostly report AIME 2026, HMMT February 2026 and IMOAnswerBench rather than FrontierMath, so check those live tables in this post and the open-weight filter on the math leaderboard. Self-hosted math models still need a verification step, because their reported scores are rarely independently reproduced.

05Can an LLM do research-level math?

Partly. The FrontierMath Tier 4 table tracks research-grade problems written by mathematicians, and the best models now solve most of them under the benchmark's conditions. That is evidence for strong problem solving, not for reliable proofs: check every derivation, and use a proof assistant when correctness matters.

Reader questions

Share or save

Share on XShare on LinkedIn

Keep readingAll research

Choose the right model before an expensive mistake. One weekly recommendation: what to choose, what costs less, and what is not worth switching to.

Join 7,000+ subscribers.