An AI model benchmark scorecard puts the published test scores of leading AI models side by side. This one compares 14 models from 7 developers, including Claude, GPT, Gemini, Grok, DeepSeek and open-weights models, on 8 benchmarks for knowledge, reasoning, coding, computer use, images and chat, with every score linked to its source.
Pick a skill area, then click any column to rank the models by it. The best score in each column is highlighted, and every score links to where the developer published it.
14 models on 8 benchmarks, highest average: Gemini 4 Carbon (4 of 6 reported)
Humanity's Last Exam (no tools)
Claude Fable 5.2
74.3%
Humanity's Last Exam (with tools)
Claude Fable 5.2
74.2%
ARC-AGI-2
GPT-6 Astra
95%
Terminal-Bench 4.0
Claude Sonnet 5.5
70.6%
OSWorld 2.1 (offline subset)
Claude Fable 5.2
89.3%
GDPval-AA
Claude Opus 5.5
1866 Elo
MMMU-Pro
Gemini 4 Carbon
89.3%
LMArena Text Elo
Gemini 4 Carbon
1526 Elo
Scores are as published by each model's developer or the benchmark's official leaderboard, checked 2026-11-14. "Vendor" scores come from the developer, "independent" ones from a third-party leaderboard; hover a score for its setting. Tests use different settings (reasoning effort, tools, harness), so small gaps are not meaningful. "-" means no published score. The average covers only the percentage benchmarks a model reports; Elo ratings are ranked on their own.
Knowledge · percent correct
Expert-written, closed-ended questions across maths, science and the humanities, answered without search or code tools. Leader here: Claude Fable 5.2 (74.3%).
Benchmark site for Humanity's Last Exam (no tools) (opens in a new tab)Knowledge · percent correct
The same expert questions, but the model may use search and code tools. Leader here: Claude Fable 5.2 (74.2%).
Benchmark site for Humanity's Last Exam (with tools) (opens in a new tab)Reasoning · percent correct
Novel visual grid puzzles that test whether a model can infer an abstract rule from a few examples. Leader here: GPT-6 Astra (95%).
Benchmark site for ARC-AGI-2 (opens in a new tab)Coding · percent correct
Hard, realistic tasks an agent must finish in a command-line terminal, checked by automated tests. Leader here: Claude Sonnet 5.5 (70.6%).
Benchmark site for Terminal-Bench 4.0 (opens in a new tab)Agentic · percent correct
An agent operates a real desktop computer through the screen, keyboard and mouse to complete everyday tasks (partial-credit score). Leader here: Claude Fable 5.2 (89.3%).
Benchmark site for OSWorld 2.1 (offline subset) (opens in a new tab)Agentic · Elo rating
Real-world knowledge-work tasks across many occupations, done agentically and compared head-to-head, reported as an Elo rating. Leader here: Claude Opus 5.5 (1866 Elo).
Benchmark site for GDPval-AA (opens in a new tab)Multimodal · percent correct
College-level questions that require reading images, charts and diagrams across many subjects. Leader here: Gemini 4 Carbon (89.3%).
Benchmark site for MMMU-Pro (opens in a new tab)Chat · Elo rating
Human preference rating from millions of blind side-by-side chat votes. Leader here: Gemini 4 Carbon (1526 Elo).
Benchmark site for LMArena Text Elo (opens in a new tab)Show all benchmarks, or only knowledge, reasoning, coding, agentic, multimodal or chat tests.
Click a column to rank models by it, or by the average of the percentage scores they report.
Narrow to one developer or to open-weights models you can run yourself.
Open any score's source, then copy a link to your view or the ranking as text.
No single model leads every test, so it depends on the job. In this scorecard (checked November 14, 2026), Claude Fable 5.2 (74.3%) leads Humanity's Last Exam, GPT-6 Astra (95%) leads ARC-AGI-2 reasoning, Claude Sonnet 5.5 (70.6%) leads Terminal-Bench coding, Claude Fable 5.2 (89.3%) leads OSWorld computer use and Gemini 4 Carbon (1526 Elo) leads LMArena chat preference.
AI benchmarks are fixed sets of tasks with automatic or expert scoring, used to compare models on the same test. Each one measures a narrow skill, such as expert knowledge (Humanity's Last Exam), abstract reasoning (ARC-AGI-2), coding in a terminal (Terminal-Bench) or operating a computer (OSWorld), so no single score tells you which model is best overall.
Developers and independent leaderboards run tests with different settings: reasoning effort, whether tools are allowed, the agent harness, the number of attempts and even the question subset. A developer's own number is often higher than an independent rerun. This scorecard marks each score as vendor or independent and shows the setting when you hover it.
They are useful but limited. Popular benchmarks saturate as models approach 100%, test questions can leak into training data, and a high score on a test does not guarantee good results on your own tasks. Treat gaps of a few points as a tie, compare several benchmarks, and try the top models on real work before you choose.
An Elo rating ranks models by head-to-head comparisons, as in chess. LMArena collects blind human votes between two chat answers, and GDPval-AA compares models on real work tasks. Elo is relative to the other models on the leaderboard, so it moves as models are added and is not a percentage.
Scores checked on November 14, 2026 from developer launch posts and model cards and from the ARC Prize, Artificial Analysis and LMArena leaderboards. AI Wins is not affiliated with any developer listed.
Compare how much text each model can read at once.
Open toolSee what the top-scoring models cost per call.
Open toolFind the best open model for your GPU or Mac.
Open toolGet matched to the right AI tool for your task.
Open toolNew models top these charts every few weeks. Get the AI breakthroughs that matter in the free daily AI Wins digest.
Get the free digestOnly positive AI stories, delivered daily or weekly. Free, no spam, unsubscribe anytime.