Kimi K3 — Benchmarks

GPQA Diamond is a set of extremely hard, graduate-level science questions (physics, chemistry, biology) written by domain experts and filtered so that skilled non-experts with web access still fail. It measures genuine reasoning rather than memorization.

LiveBench Reasoning measures logical, multi-step reasoning using contamination-free questions that are refreshed regularly, so models cannot have trained on the test set.

ARC-AGI-2 is the second-generation ARC Prize puzzle set, rebuilt so that tasks solved by brute-force search or memorization were removed and every problem still requires inferring a novel rule from a handful of examples. It is markedly harder than the original ARC-AGI, and most models — open and closed alike — score near zero.

ARC-AGI tests fluid, abstract reasoning on small visual grid puzzles where each task follows a novel rule the model must infer from a few examples. It deliberately resists memorization and is one of the most-watched measures of general reasoning progress.

SimpleBench is a set of everyday, common-sense and trick questions that humans answer easily but language models often get wrong. It probes basic reasoning and robustness rather than specialist knowledge.

APEX-Agents (AI Productivity Index) is Mercor's evaluation of AI agents on long-horizon, multi-application professional-services work — investment-banking, management-consulting, and corporate-law style tasks — graded against expert-produced solutions. The score is Pass@1, the fraction solved correctly on the first attempt.

Chess Puzzles is an Epoch AI evaluation that asks models to find the winning move in tactical chess positions drawn from real games. Because each puzzle has one verifiable correct answer and no natural-language shortcut, it isolates concrete multi-step planning.

CritPt (Critical Points) tests graduate-level physics problem-solving on research-style questions spanning mechanics, electromagnetism, quantum mechanics, and more. It was created by physicists to probe whether a model can reason through multi-step physics rather than pattern-match textbook answers.

DTBench is a set of handcrafted, expert-validated multiple-choice questions on decision theory — Newcomb-like problems and related puzzles — built by the Conceptual Reasoning Index project with Anthropic. It tests whether a model can follow an abstract argument rather than recall a fact.

ForecastBench, from the Forecasting Research Institute, asks models to predict real-world future events, with questions that only resolve after the model's training cutoff so the answers can't have been memorised. It is scored as a difficulty-adjusted Brier index, where higher is better.

Furniture Assembly, built by Epoch AI, tests spatial reasoning with 60 photos of partly built IKEA furniture and the assembly manual: the model must say whether the build is correct so far and, if not, name the mistaken step and describe the error. Scores are accuracy across the photos, and a human baseline averaged 63%.

LiveBench 2026 Reasoning measures logical, multi-step problem solving. It belongs to the 2026 LiveBench release (first snapshot September 2, 2026), a redesigned LiveBench in which each of seven categories is scored by a single task and questions are refreshed to limit training-set contamination. It is a different test from the original LiveBench boards, so the scores are not comparable.

LMCA (Language Model Conceptual Argumentation) scores models on expert-rated arguments across philosophy, decision theory, and AI risk — domains where there is no empirical feedback loop to check an answer against. It is part of the Conceptual Reasoning Index, built with Anthropic by researchers Emery Cooper and Caspar Oesterheld.

Mystery Game Puzzles is an Epoch AI evaluation of multi-step deductive reasoning: the model works through a mystery or logic-game scenario to arrive at the correct solution. It probes sustained, structured inference rather than single-step question answering.

Vending-Bench 2, from Andon Labs, is a long-horizon agent benchmark in which the model runs a simulated vending-machine business for a year, handling suppliers, pricing and cash flow. The score is the final money balance in US dollars.

LiveBench Coding evaluates code generation and completion on fresh, contamination-free programming tasks that are updated regularly.

SciCode asks a model to write real scientific-computing code — implementing multi-step numerical methods drawn from physics, chemistry, biology, and math research — and checks the output against reference solutions. It was built by a team of scientists and ML researchers to test coding ability on genuine research problems, not typical software tasks.

WeirdML asks models to write machine-learning code for small, deliberately unusual datasets, then runs the code and scores the accuracy it actually achieves. Built by Håvard Tveit Ihle, it rewards practical problem-solving over recognizing a familiar task.

DeepSWE, built by Datacurve, evaluates coding agents on 113 original software-engineering tasks across 91 repositories and five languages — written from scratch rather than adapted from real commits, so a model can't have seen the solution during pretraining, and graded Pass@1 with hand-written behavioral verifiers. It's a benchmark name only, unrelated to the open-weight DeepSWE-Preview coding model.

FrontierCode, from Cognition (the team behind Devin) with maintainers of open-source projects, gives an agent a real repository and issue and grades its patch on whether it's mergeable — held-out tests plus a maintainer-written rubric covering correctness, regression safety, style and project conventions. The Main set is the 100 hardest tasks.

ALE-Bench, from Sakana AI and AtCoder, tests long-horizon algorithmic optimisation using problems drawn from AtCoder Heuristic Contests, where there is no single correct answer and solutions are scored on quality. Results are reported as an AtCoder-style performance rating, where higher is better.

FrontierSWE, from Proximal, tests coding agents on long-horizon software-engineering tasks spanning implementation, performance optimisation and open-ended research, each run for hours through an agent harness. The score is the average normalised task score across repeated runs of every task.

LiveBench 2026 Coding measures whether a model writes correct code for programming problems. It belongs to the 2026 LiveBench release (first snapshot September 2, 2026), a redesigned LiveBench in which each of seven categories is scored by a single task and questions are refreshed to limit training-set contamination. It is a different test from the original LiveBench boards, so the scores are not comparable.

LiveBench 2026 Long-Horizon measures multi-step tasks that a model has to carry through many steps to finish. It belongs to the 2026 LiveBench release (first snapshot September 2, 2026), a redesigned LiveBench in which each of seven categories is scored by a single task and questions are refreshed to limit training-set contamination. It is a different test from the original LiveBench boards, so the scores are not comparable.

Surface Evolver Bench tests whether a model can write a correct simulation file for Surface Evolver, a physics tool that models surfaces shaped by tension, gravity and volume constraints, defining a 3D structure in its domain-specific format from only a few examples. Scores compare the resulting geometry — volume, area and energy — against reference solutions; the benchmark was built by an independent developer (GitHub: yhenon).

AIME (American Invitational Mathematics Examination) is a prestigious high-school competition of hard, integer-answer problems. It's a widely-cited yardstick for multi-step mathematical reasoning, here on the 2024–2025 papers.

FrontierMath is a benchmark of exceptionally hard, original research-level mathematics problems created with professional mathematicians. Even the strongest models solve only a small fraction, making it a frontier measure of genuine mathematical ability.

LiveBench Math measures mathematical problem-solving on contamination-free, regularly-refreshed questions, including competition-style problems.

FrontierMath Tier 4 is the hardest tier of Epoch AI's research-level mathematics benchmark, holding problems that working mathematicians consider genuinely difficult research questions. Scores here stay low even for frontier models, making it the sharpest available read on mathematical ability.

ProofBench evaluates a model's ability to write and check rigorous mathematical proofs, rather than just produce a final numeric answer. It probes step-by-step logical correctness, which is a harder and more distinct skill than solving for a number.

SimpleQA measures factual accuracy on short, fact-seeking questions with a single correct answer — directly probing how often a model is right versus confidently wrong (hallucination) on simple facts.

Arena (arena.ai, formerly LMArena / Chatbot Arena) ranks models by millions of blind, head-to-head human votes on real prompts. The Arena Score shown here is a Bradley-Terry rating computed with Arena's default style control, so higher means people preferred that model's answers more often — it's a relative rating, not a percentage, so compare it only to other models on this board.