AIME-style math benchmarks are now so saturated at the frontier that leaderboards disagree about who is even in first place — and that disagreement, not the scores themselves, is the real signal.
The American Invitational Mathematics Examination (AIME) was adopted as an LLM benchmark because it was hard, short, and had a clean numeric answer that’s easy to grade automatically. In 2026 it has stopped doing the job it was hired for. Multiple independent trackers now put frontier models within a few points of a perfect score, and — more tellingly — they don’t agree on which model that is.
What the AIME 2026 trackers actually show
Three trackers, three different leaderboards, run close together in time:
- BenchLM’s AIME26 snapshot (dated July 23, 2026) puts GLM-5.2 first at 99.2%, with Beam at 97.8% and Inkling at 97.1%; Kimi K2.6 trails at 96.4%. BenchLM’s own read is that the top three sitting within 2.8 points of each other is itself evidence the benchmark is saturating for frontier models.
- llm-stats.com’s AIME 2026 page ranks Kimi K2.6 first — at 96.4%, the same number BenchLM has it trailing at — with an average across all tracked models of 83.8%. The page itself discloses a reliability caveat: 0 of its results are independently verified, versus 12 self-reported by vendors.
- A Developers Digest board, sourced from MathArena, has GPT-5.5 (xhigh reasoning effort) and Claude Opus 4.8 (max effort) tied at 100.00% as of October 17, 2026 — but also notes that MathArena has since marked its AIME 2026 track as deprecated, which suggests even the benchmark’s maintainers think it has stopped discriminating between models.
Put plainly: ask three trackers who leads AIME 2026 and you get three different winners, on three different scales, with no shared ground truth. That is what a dying benchmark looks like from the outside.
Why competition math saturates fast
AIME problems are drawn from a finite, well-studied genre — integer answers, a few dozen recurring problem archetypes, and (crucially) decades of solved past papers that are almost certainly in pretraining data in some form. A benchmark built from a closed, high-school/early-college competition format has a ceiling that modern reasoning models can simply reach. This is the same contamination-and-saturation story documented elsewhere on reasoning benchmarks — see our Humanity’s Last Exam coverage for how a benchmark explicitly designed to resist this problem is holding up, and our FrontierMath breakdown for what happened when Epoch AI had to rewrite 42% of its own problem set after contamination concerns.
| Benchmark | Status in 2026 | Why |
|---|---|---|
| AIME 2026 | Saturating / deprecated by at least one tracker (MathArena) | Closed problem genre, top models within ~3 points of ceiling, multiple 100% ties reported |
| MATH-500 | Still shows separation per BenchLM | Broader problem set, higher ceiling difficulty mix |
| BRUMO 2025 (Bulgarian Mathematical Olympiad) | Still shows separation per BenchLM | Less-studied competition corpus, lower contamination risk |
| FrontierMath | Actively maintained, researcher-graded | Novel, unpublished problems; Epoch AI rewrote 42% after contamination concerns |
What’s replacing contest math as a signal
The field’s response has been to move toward benchmarks that are harder to memorize because the problems are new, expert-graded, or deliberately obscure: FrontierMath’s unpublished problem sets, Humanity’s Last Exam’s expert-submitted questions across disciplines, and narrower community sets like BRUMO 2025 that haven’t been scraped to death yet. The pattern mirrors what happened to HumanEval in code generation and the original MMLU — a benchmark proves useful, gets gamed or memorized, and the field routes around it rather than fixing it. See our guide to LLM evaluation methodology for how to weigh a saturating benchmark against newer alternatives when comparing models.
The reliability problem underneath the scores
Beyond saturation, there’s a measurement problem: llm-stats.com’s own disclosure that it has zero independently verified AIME 2026 results, relying instead on vendor self-reports, is a reminder that a lot of “benchmark” numbers circulating for any given model are press-release numbers, not reproduced ones. Cross-checking a claimed score against at least two trackers — and treating disagreement as a yellow flag rather than noise — is standard practice worth applying here.
How vendors turn a saturated benchmark into a marketing number
A saturated benchmark doesn’t stop being cited — it just becomes a worse instrument while remaining a popular one, because “100% on AIME” is still a clean headline even after it has stopped meaning much. Watch for three patterns when a vendor cites an AIME 2026 score in a release post: whether the number is self-reported or independently reproduced by a third party, whether the reasoning-effort setting used to get that score is disclosed (the Developers Digest board’s 100% ties were both reported at maximum reasoning effort, which costs far more in tokens than a default setting — see our coverage of how hallucination rates are measured for a parallel case where effort settings change the reported number), and whether the comparison set includes models that have already saturated the test versus only cherry-picked competitors that haven’t.
None of this means AIME-style benchmarks were a bad idea. For roughly two years they were one of the few automatically-gradable signals that correlated reasonably well with a model’s general quantitative reasoning ability, and they’re cheap to run compared to expert-graded alternatives like FrontierMath or Humanity’s Last Exam, which require domain specialists to write and check problems. The honest conclusion is closer to a lifecycle: a benchmark is useful while it discriminates between models, and AIME 2026 has, by the admission of at least one of its own maintainers, stopped doing that at the frontier even though it may still usefully separate mid-tier and budget models from frontier ones.
What to check before trusting any AIME 2026 claim
Given the spread between trackers documented above — a 96.4% that’s simultaneously someone’s “first place” and someone else’s “fourth place,” and a 100% that one source reports and another source says is being retired — a reader comparing models on this benchmark in late 2026 should treat any single-tracker number as provisional. Checking at least two of the three trackers cited here, noting the reported reasoning-effort setting, and checking the snapshot date (leaderboards from different months are not directly comparable once vendors ship mid-cycle model updates) are the minimum due-diligence steps before repeating a vendor’s AIME claim as fact.
FAQ
Is AIME still a useful way to compare frontier models in 2026?
Not on its own. With top models clustered within a few points of 100% across most trackers, and at least one maintainer (MathArena) deprecating its AIME 2026 track, the benchmark no longer discriminates well at the frontier. It remains useful as a sanity check for smaller or older models that haven’t saturated it yet.
Why do different trackers report different AIME 2026 leaders?
Different snapshot dates, different model sets, different grading harnesses (self-reported vendor numbers vs. independently run evaluations), and in at least one case a different underlying source (MathArena vs. a tracker’s own run). None of the three trackers reviewed here agree on both the leader and the score.
What benchmarks are replacing AIME for math reasoning?
FrontierMath and Humanity’s Last Exam for expert-graded, harder-to-memorize problems; BRUMO 2025 and MATH-500 as contest-style benchmarks that still show meaningful separation between models, per BenchLM’s own comparisons.
Last updated 2026-11-13. This page is refreshed as benchmarks and scores move.