MathematicsSaturated

AIME

Also known as American Invitational Mathematics Examination.

Hard high-school competition math with exact integer answers.

  • 229 models have a score, from Epoch AI.
  • The leader is gemini-4-argon_high at 100%, dated Oct 13, 2026.
  • The top 10 are within 0 points of each other, close enough that most differences there are noise.
  • Its newest result is dated Nov 6, 2026, and 16 results were published in the last 30 days.
  • A score on this page last changed on Nov 11, 2026.

Metric

accuracy (%)

Status

Saturated

Category

Mathematics

Maintainer

MAA; curated by Epoch AI

Chance baseline

0%

What it measures

The American Invitational Mathematics Examination (AIME) measures advanced pre-college mathematical problem solving: multi-step, creative reasoning across algebra, geometry, number theory, and combinatorics. It is a human competition first, organized by the Mathematical Association of America (MAA) for students who have already excelled on the AMC 10 or AMC 12. As a language-model benchmark it has become one of the standard tests of mathematical reasoning, because its problems demand a genuine multi-step derivation rather than fact recall, and because each problem has a single integer answer that is trivial to grade automatically.

The task in detail

Each AIME is a 15-question, 3-hour test. Every answer is an integer between 000 and 999 inclusive, and no answer choices are provided, so a solver must actually produce the number. Calculators, rulers, protractors, and other aids are not allowed. Problems cover elementary algebra, geometry, trigonometry, number theory, probability, and combinatorics, and they escalate in difficulty through the paper, from approachable early questions to hard closing ones that stump most qualifiers. Between 2000 and 2026 the MAA offered two forms each year, AIME I and AIME II (the second an alternate for students who could not sit the first), of equal standing. The AIME sits between the AMC and the USA Mathematical Olympiad (USAMO or USAJMO) in the qualification chain.

For language models the exam is used almost verbatim: the model is given a problem statement and must output the integer answer, usually after chain-of-thought working. Because a full AIME is only 15 problems, and only about 30 fresh problems appear per year, evaluators either aggregate several years, run each problem many times, or use held-out "mock" sets. Epoch AI, for example, hosts a Mock AIME 2024-2025 assembled from 45 problems across three OTIS mock exams (15 from Mock AIME 2024, 15 from Mock AIME 2025 I, and 15 from Mock AIME 2025 II), positioned as harder than the older MATH Level 5 benchmark but easier than research-level sets.

How models are scored

The dominant metric is exact-match accuracy, often reported as pass@1: the fraction of problems whose final integer answer exactly matches the key, after the answer is extracted from the model's output. There is no partial credit, mirroring the human contest, where one point is earned per correct answer, no points are deducted for wrong ones, and scores are integers from 0 to 15.

Because a single run of a probabilistic model is noisy on so few problems, careful evaluators sample each problem many times and average. Epoch AI runs most models 16 times on its Mock AIME set, uses each model's API-default temperature, zero-shot chain-of-thought prompting, and an empty system prompt, and reports a mean with a confidence interval of plus or minus one standard error. Other groups report consensus (majority-vote) accuracy or best-of-n, which raise the number but measure something different from a single attempt, so the sampling method always has to be read alongside the score.

How to read a score

Always check three things: which year's problems, how many samples, and whether the number is pass@1 or a consensus or majority-vote figure. A "90% on AIME" from majority voting over dozens of samples is not comparable to 90% from a single attempt. Because there are only 15 problems per exam, one problem is worth nearly 7 percentage points, so single-run scores on one year are statistically shaky and confidence intervals matter.

Contamination is the other lens. If a model may have seen a given year's problems and solutions during training, a high score can reflect memorization rather than reasoning, which is why fresh exams and mock sets are used as a sanity check. Treat a high AIME number as strong evidence of reasoning mainly when it holds up on problems released after the model's training cutoff.

History and notable results

The AIME has been a fixture of American math competitions for decades, but it became a marquee AI benchmark in 2024. For most of the language-model era it was a hard target: general chat models solved only a small fraction of problems. That changed sharply with the arrival of dedicated reasoning models. OpenAI presented its first o1 model in late 2024 with a large reported jump on AIME 2024 over the previous GPT-4o generation, and the result was widely cited as a turning point for test-time reasoning.

Through 2025, frontier reasoning systems pushed AIME accuracy to very high levels, to the point that Epoch AI notes older math benchmarks are approaching saturation and added a harder Mock AIME set to keep resolving differences between top models. Because the standings move with every release, the live leaderboard is the place to see current results. The exam itself is not static either: the two-form structure of AIME I and AIME II ran from 2000 through 2026, after which the MAA is changing the exam's format.

Limitations and criticism

The biggest issue is data contamination: past AIME problems and their solutions are all over the web, so a model may have effectively memorized them. The tiny size (15 problems per form) makes single-run scores noisy and easy to over-interpret. The integer-answer format means a model can occasionally reach the right number through a flawed or lucky route, which exact-match scoring cannot detect. And scores are sensitive to prompt and sampling choices, so headline numbers from different labs are often not directly comparable.

Epoch AI's Mock AIME 2024-2025 is the closest sibling, a harder, less-contaminated variant built from mock exams. The MATH dataset and its hardest MATH Level 5 tier draw on AIME and AMC problems for a broader test of the same skills. Above AIME sit olympiad-level challenges built from USAMO and IMO problems, along with research-grade sets such as FrontierMath, while general reasoning suites like GPQA Diamond cover scientific rather than purely mathematical reasoning.

References

Leaderboard

Checked Nov 11, 2026, 07:00 UTC · scores last changed Nov 11, 2026

Best AIME score per model, ranked. Each score links to its source.

#ModelScoreSource
1Google: Gemini 4 Argon(high)Google DeepMind100%epoch
2gpt-5.5-pre-release_xhigh(xhigh)OpenAI100%epoch
3gpt-5.5-pro-pre-release_xhigh(xhigh)OpenAI100%epoch
4Qwen: Qwen3.8 Max (0902)(xhigh)Alibaba100%epoch
5Anthropic: Claude Fable 5.1(max)Anthropic100%epoch
6OpenAI: GPT-6.1 Terra(max)OpenAI100%epoch
7Anthropic: Claude Opus 5.5(max)Anthropic100%epoch
8OpenAI: GPT-6.1 Sol(max)OpenAI100%epoch
9Anthropic: Claude Sonnet 5.5(max)Anthropic100%epoch
10OpenAI: GPT-6 Sol(max)OpenAI100%epoch
11OpenAI: GPT-6 Astra(max)OpenAI100%epoch
12gpt-5.5-pre-release_xhigh(xhigh)OpenAI100%epoch
13gpt-5.5-pro-pre-release_xhigh(xhigh)OpenAI100%epoch
14kimi-k3.1-pro-together_max(max)Moonshot100%epoch
15Google: Gemini 4 Carbon(high)Google DeepMind100%epoch
16Anthropic: Claude Fable 5.2(max)Anthropic99.7%epoch
17qwen3.8-max_xhigh(xhigh)Alibaba99.4%epoch
18muse-wonder-1.0_max(max)Meta AI99.4%epoch
19mistral-large-4-0-oss_high(high)Mistral AI99.4%epoch
20DeepSeek: DeepSeek V4.1 Flash 1017(max)DeepSeek99.4%epoch
21doubao-seed-evolving_high(high)ByteDance99.2%epoch
22Meta: Muse Spark 1.3(xhigh)Meta AI99.2%epoch
23SpaceXAI: Grok 4.6(xhigh)xAI99.2%epoch
24Google: Gemini 3.8 Flash(high)Google DeepMind98.9%epoch
25Anthropic: Claude Haiku 5.5(xhigh)Anthropic98.9%epoch
26OpenAI: GPT-6 Luna(max)OpenAI98.9%epoch
27DeepSeek: DeepSeek V4 Pro 0813(max)DeepSeek98.6%epoch
28DeepSeek: DeepSeek V4.1 Flash(max)DeepSeek98.3%epoch
29Mistral: Mistral Large 4(high)Mistral AI98.3%epoch
30SpaceXAI: Grok 4.7(xhigh)xAI98.1%epoch
31Meta: Muse Spark Open(xhigh)Meta AI98.1%epoch
32SpaceXAI: Grok 4.5(high)xAI97.8%epoch
33Anthropic: Claude Fable 5(low)Anthropic97.8%epoch
34Google: Gemini 3.7 Flash(high)Google DeepMind97.2%epoch
35doubao-seed-2.1-turbo-260628_high(high)ByteDance96.1%epoch
36MoonshotAI: Kimi K2.6Moonshot96.1%epoch
37doubao-seed-2.1-pro-260628_high(high)ByteDance95.6%epoch
38Google: Gemini 3 Flash Preview(high)Google DeepMind95.6%epoch
39Google: Gemini 3.1 Pro Preview(high)Google DeepMind95.6%epoch
40MoonshotAI: Kimi K2.7 CodeMoonshot95.6%epoch
41OpenAI: GPT-5.4(medium)OpenAI95.6%epoch
42Qwen: Qwen3.7 MaxAlibaba95.6%epoch
43claude-opus-4-6_64KAnthropic94.4%epoch
44DeepSeek: DeepSeek V4 Flash 0731(max)DeepSeek94.4%epoch
45doubao-seed-2.0-pro-260215_high(high)ByteDance93.9%epoch
46Z.ai: GLM 5.3 Flash(max)Z.ai (Zhipu AI)93.9%epoch
47Z.ai: GLM 5.1Z.ai (Zhipu AI)93.3%epoch
48Qwen: Qwen3.7 PlusAlibaba93.3%epoch
49SpaceXAI: Grok 4.3(high)xAI93.3%epoch
50Qwen: Qwen3.6 PlusAlibaba93.3%epoch
51Anthropic: Claude Opus 5(low)Anthropic93.3%epoch
52claude-opus-4-6_32KAnthropic93.1%epoch
53grok-4.20-0309-reasoningxAI92.2%epoch
54grok-4.20-0309-reasoningxAI92.2%epoch
55fireworks/kimi-k2p5Moonshot92.2%epoch
56fireworks/kimi-k2p5Moonshot92.2%epoch
57gemini-3-pro-previewGoogle DeepMind91.4%epoch
58gemini-3-pro-previewGoogle DeepMind91.4%epoch
59Anthropic: Claude Opus 4.6(max)Anthropic91.1%epoch
60qwen3.6-max-previewAlibaba91.1%epoch
61qwen3.6-max-previewAlibaba91.1%epoch
62doubao-seed-2.0-lite-260428_high(high)ByteDance91.1%epoch
63Z.ai: GLM 5.3(max)Z.ai (Zhipu AI)91.1%epoch
64Qwen: Qwen3.6 27BAlibaba91.1%epoch
65Thinking Machines: Inkling Small(xhigh)Thinking Machines90%epoch
66muse-sparkMeta AI88.9%epoch
67muse-sparkMeta AI88.9%epoch
68OpenAI: GPT-5.4 Mini(xhigh)OpenAI88.9%epoch
69Qwen: Qwen3.5 397B A17BAlibaba88.9%epoch
70Google: Gemini 3.5 Flash(low)Google DeepMind88.9%epoch
71Thinking Machines: Inkling(xhigh)Thinking Machines88.9%epoch
72OpenAI: gpt-oss-120b(high)OpenAI88.9%epoch
73deepseek-reasonerDeepSeek87.8%epoch
74deepseek-reasonerDeepSeek87.8%epoch
75qwen3.5-plusAlibaba86.7%epoch
76Qwen: Qwen3 235B A22B Thinking 2507Alibaba86.7%epoch
77Qwen: Qwen3.6 35B A3BAlibaba86.7%epoch
78Qwen: Qwen3.7 FlashAlibaba86.7%epoch
79Anthropic: Claude Opus 4.7(max)Anthropic86.7%epoch
80NVIDIA: Nemotron 3 UltraNvidia86.7%epoch
81claude-opus-4-5-20251101_32KAnthropic86.1%epoch
82claude-sonnet-4-6_32KAnthropic85.8%epoch
83Google: Gemini 2.5 ProGoogle DeepMind84.7%epoch
84Qwen: Qwen3.6 FlashAlibaba84.4%epoch
85OpenAI: o3(medium)OpenAI84.4%epoch
86qwen3.5-flashAlibaba84.4%epoch
87Anthropic: Claude Opus 4.8(none)Anthropic84.4%epoch
88grok-4-0709xAI84%epoch
89grok-4-0709xAI84%epoch
90Z.ai: GLM 4.7Z.ai (Zhipu AI)83.3%epoch
91kimi-k2-thinking-turboMoonshot83.1%epoch
92kimi-k2-thinking-turboMoonshot83.1%epoch
93Google: Gemma 4 26B A4B (minimal)Google DeepMind82.2%epoch
94claude-opus-4-5-20251101_16KAnthropic81.7%epoch
95Z.ai: GLM 5Z.ai (Zhipu AI)80%epoch
96Google: Gemini 3.6 Flash(minimal)Google DeepMind80%epoch
97Anthropic: Claude Sonnet 5(max)Anthropic80%epoch
98grok-3-mini-beta_high(high)xAI77.8%epoch
99claude-sonnet-4-5-20250929_32KAnthropic77.8%epoch
100claude-sonnet-4-5-20250929_59KAnthropic77.8%epoch

Run your own prompt on the top 3

Gemini 4 Argon, Qwen3.8 Max (0902), Claude Fable 5.1, side by side on the same prompt, with each one's time and cost. A free account includes starter credit.

Compare them

Compare the leaders

More Mathematics benchmarks

AIME is maintained by MAA; curated by Epoch AI. Scores on this page are collected from openly-licensed sources and each links to its origin. See the full matrix to compare every model across every benchmark.