{
 "body": "\n## What it measures\n\nAIME 2026 repurposes the 30 problems from the 2026 American Invitational Mathematics Examination as a language-model reasoning test. Like every AIME sitting, it sits above the AMC 10/12 in the US math-competition pipeline: only students who clear a qualifying AMC score are invited to sit it, so the problems assume real fluency in algebra, geometry, number theory and combinatorics rather than rote recall. Each problem has exactly one correct integer answer between 0 and 999, so there is no answer key to pattern-match against and no partial credit for a nearly-right derivation. AIME I 2026 was administered February 5, 2026 and AIME II 2026 on February 11, 2026, both confirmed by the MAA's own announcement and by problems and answer keys already published on the Art of Problem Solving wiki - this is a completed, scored competition as of this page's research date, not a future one.\n\n## How it is scored\n\nScoring is accuracy: the fraction of the 30 problems for which the model's final integer matches the official answer, most often reported as pass@1 for a single sampled attempt, though some reporters average several samples per problem or take a majority vote (cons@k) instead, which can move a score noticeably on a set this small. Because answers are free-response integers rather than multiple-choice letters, there is no meaningfully citable random baseline, and tool access (for example, whether a model may run code to check arithmetic) is part of the evaluation condition rather than a footnote.\n\n## Dataset and licence\n\nThe 30 problems split evenly between AIME I 2026 and AIME II 2026, both administered in February 2026. As with earlier AIME years, the MAA does not publish an official machine-readable eval package, so the problems have been independently transcribed by multiple groups: `math-ai/aime26`, created by Yifan Zhang and the Math-AI team and last updated in February 2026, states an Apache-2.0 licence for its transcription; MathArena maintains a parallel `MathArena/aime_2026` dataset for its own tracking. As with the earlier sittings in this series, these licences cover each group's transcription work, not the underlying exam, which the MAA itself owns.\n\n## Who publishes it\n\nThe exam is written and administered by the Mathematical Association of America. As a language-model benchmark, no single lab or paper owns AIME 2026; it is packaged and run independently by several groups, including math-ai (the dataset inspect_evals loads), MathArena (which lists AIME 2026 among the competitions on its live leaderboard), and OpenCompass (`aime_2026`). Model developers are expected to report it themselves in system cards and technical reports, following the same pattern as the earlier sittings in this series.\n\n## Lineage\n\nAIME 2026 is the most recent sitting in this repository's AIME series, following `aime_2025` and, before that, `aime_2024`. Note the id spelling: this repository's existing `aime_2025` page uses an underscore before the year, while this page and `aime_2024` do not; all three refer to the same annual AIME series under one consistent format, and the inconsistency is a repository data-entry artefact rather than a meaningful distinction, recorded in `benchmarks/_census/DATA-QUALITY.md`. All three years share an identical problem format and scoring approach and differ only in which year's 30 problems they use.\n\n## Saturation and contamination\n\nMathArena lists AIME 2026 among the competitions its live leaderboard tracks, but that leaderboard renders its tables client-side, and no specific model score could be extracted from a page fetch during this research, so a top score is not established here. As context rather than a substitute figure: the immediately preceding sitting, AIME 2025, was reported by one tracker at roughly 94.6% for a top model about six months after its own sitting and was dropped from Artificial Analysis's Intelligence Index entirely in a January 2026 refresh for being saturated - a pattern that, if it repeats, would suggest AIME 2026 may already be approaching a similar ceiling seven months after its own sitting, but this is inference from a sibling benchmark's trajectory, not a reading of AIME 2026 itself. Contamination risk is medium: the problems and answer key have been public for about seven months by this page's research date, noticeably less exposure than AIME 2024 (over two years) or AIME 2025 (roughly a year and a half), but long enough that models trained after February 2026 could plausibly have encountered them.\n\n## How to run it\n\ninspect_evals exposes it as the `aime_2026` task, loading the `math-ai/aime26` dataset and grading by extracting a final boxed or stated integer and comparing it to the answer key; OpenCompass exposes its own `aime_2026` dataset configuration. Both are free-response tasks, so scores are sensitive to prompt format, to whether code execution is permitted, and to sample count, given there are only 30 problems in total and a single pass@1 run carries meaningful sampling noise.\n\n## Reading the numbers\n\nBecause AIME 2026 is the newest sitting in this series as of this page's research date, a high score here is currently better evidence of genuine competition-math ability than the same score on AIME 2024 or AIME 2025 would be, simply because it has had far less time to enter training data. That advantage is temporary: based on how quickly the two preceding AIME years were reported as saturated, expect AIME 2026 to follow the same path within roughly a year of its sitting. Until a top score is confirmed from a live source, treat any claim about AIME 2026 being \"solved\" or \"saturated\" as unverified, and check whether a reported number used tools, multiple samples, or a single pass@1 draw before comparing two models' figures.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "aime_2026",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "aliases": [
   "AIME26",
   "AIME 2026 I and II"
  ],
  "category": "math",
  "contamination": {
   "note": "AIME 2026's problems and official answer key have been public since shortly after each February 2026 sitting - about seven months by this page's research date - considerably less exposure time than AIME 2024 (over two years) or AIME 2025 (about a year and a half), but still enough that a model with a training-data cutoff after February 2026 could plausibly have seen these exact problems.",
   "risk": "medium"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "Apache-2.0",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 30,
   "size_note": "30 problems total: 15 from AIME I 2026, administered February 5, 2026, and 15 from AIME II 2026, administered February 11, 2026 (dates per the MAA's own December 2025 announcement of that year's AIME thresholds).",
   "splits": "single test split, 30 rows; no train or validation split",
   "url": "https://huggingface.co/datasets/math-ai/aime26"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "sonnet-5 agent, batch 3, slice C",
   "reviewed": "",
   "reviewed_by": ""
  },
  "harness": {
   "bigbench": "",
   "helm": "",
   "inspect_evals": "aime_2026",
   "lm_eval": "",
   "opencompass": "aime_2026",
   "other": ""
  },
  "id": "aime_2026",
  "last_updated": "",
  "leaderboard_url": "https://matharena.ai/",
  "lineage": {
   "family": "",
   "predecessor": "aime_2025",
   "successors": [],
   "variants": []
  },
  "measures": "AIME 2026 gives a model the problems from the 2026 American Invitational Mathematics Examination, a competition only the top-scoring AMC 10/12 participants are invited to sit, and checks whether the model returns the single correct integer answer. It exercises multi-step algebra, geometry, number theory and combinatorics reasoning well above grade-school math benchmarks, and gives no partial credit for a sound method that lands on the wrong final number. Both 2026 sittings had already taken place and had their problems and answer keys published well before this page's research date of 2026-09-08, so this page documents a completed, scored competition rather than a forward-looking one.",
  "metric": {
   "baseline_note": "No source read for this page gave a specific model score on AIME 2026 at any sampling protocol; see Saturation and contamination for what could and could not be confirmed.",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 100,
   "name": "accuracy (pass@1, exact match on the final integer)",
   "random_baseline": null,
   "unit": "%"
  },
  "name": "AIME 2026",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "",
   "title": "",
   "url": "",
   "year": null
  },
  "publisher": {
   "authors": [],
   "org": "Mathematical Association of America (MAA)",
   "url": "https://maa.org/maa-invitational-competitions/"
  },
  "released": "2026-02",
  "repo_url": "",
  "saturation": {
   "as_of": "",
   "note": "MathArena lists 'AIME 2026' among its currently tracked competitions (confirmed via its own competition-selector list and its eth-sri/matharena GitHub config, aime_2026.yaml), but its leaderboard tables render client-side and no specific model score could be extracted from a page fetch during this research, so a top score is not established here. For context, not a substitute: AIME 2025, the immediately preceding sitting, was reported by one tracker at roughly 94.6% for a top model about six months after its sitting, and was dropped entirely from Artificial Analysis's Intelligence Index in a January 2026 refresh for being saturated, roughly eleven months after its own sitting - suggesting AIME 2026 could plausibly be following a similar trajectory by this page's research date, seven months on, but this is inference from a sibling benchmark, not a reading of AIME 2026 itself.",
   "status": "unknown",
   "top_score": null
  },
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "MAA Invitational Competitions",
    "url": "https://maa.org/maa-invitational-competitions/"
   },
   {
    "accessed": "2026-09-08",
    "title": "2025-26 AIME Thresholds Are Now Available - MAA (confirms February 5 and 11, 2026 sitting dates)",
    "url": "https://maa.org/news/2025-26-aime-thresholds-are-now-available/"
   },
   {
    "accessed": "2026-09-08",
    "title": "math-ai/aime26 dataset card",
    "url": "https://huggingface.co/datasets/math-ai/aime26"
   },
   {
    "accessed": "2026-09-08",
    "title": "2026 AIME I, Art of Problem Solving wiki (confirms February 5, 2026 administration, problems and answer key published)",
    "url": "https://artofproblemsolving.com/wiki/index.php/2026_AIME_I"
   },
   {
    "accessed": "2026-09-08",
    "title": "inspect_evals aime_2026 task (loads math-ai/aime26)",
    "url": "https://github.com/UKGovernmentBEIS/inspect_evals/tree/main/src/inspect_evals/aime_2026"
   },
   {
    "accessed": "2026-09-08",
    "title": "OpenCompass dataset configs (aime_2026 directory confirmed present)",
    "url": "https://github.com/open-compass/opencompass/tree/main/opencompass/configs/datasets"
   },
   {
    "accessed": "2026-09-08",
    "title": "MathArena live leaderboard (AIME 2026 listed among tracked competitions)",
    "url": "https://matharena.ai/"
   },
   {
    "accessed": "2026-09-08",
    "title": "MathArena aime competition configs (aime_2026.yaml)",
    "url": "https://github.com/eth-sri/matharena/tree/main/configs/competitions/aime"
   }
  ],
  "status": "active",
  "subcategory": "competition mathematics",
  "summary": "The 30 problems from the 2026 American Invitational Mathematics Examination, administered in February 2026, scored as an exact-answer test of competition math reasoning.",
  "tags": [
   "math",
   "competition-math",
   "reasoning",
   "exact-match"
  ],
  "task_format": "Free-response competition math problem in, single integer answer from 0 to 999 out; no answer choices are offered."
 }
}