What the numbers actually measure.

A composite of ten evaluations that Artificial Analysis runs itself, covering agentic work (GDPval-AA, AutomationBench, AA-Briefcase, GDP.pdf), coding (Terminal-Bench 4.0, SciCode), and knowledge and reasoning (HLE, CritPt, AA-Omniscience, AA-LCR).

Why it matters. It is the closest thing to a neutral overall grade. Every model runs the same harness, so vendor tuning does not leak in.

Watch out. The index is re-versioned as evaluations saturate, so compare scores only within the same version. We track v4.3.

1. Claude Opus 5.558
2. Muse Wonder Full56
3. Claude Sonnet 5.556
4. Claude Fable 5.153
5. Gemini 4 Argon53

Resolve real issues from large, actively maintained repositories, some of them private, and pass the hidden tests.

Why it matters. It is the current standard for "can this model fix a real bug in a real codebase", and harder to memorize than its predecessor.

Watch out. Scores depend heavily on the agent scaffold. Vendor numbers usually use the vendor’s own harness.

1. Claude Opus 5.589.9%^v
2. Claude Sonnet 5.581.3%^v
3. Claude Fable 5.181.2%^v
4. Claude Fable 580%^v
5. Claude Opus 579.2%^v

Resolve 500 human-validated GitHub issues from popular Python projects.

Why it matters. The long-running coding yardstick, useful for comparing against older models.

Watch out. Near saturation at the frontier, and the public repositories raise contamination risk. Prefer SWE-bench Pro for new models.

1. Claude Sonnet 585.2%^v
2. Beam80.9%^v
3. Gemini 3.1 Pro Preview80.6%^v
4. Inkling-Small80.2%^v
5. Tencent Hy378%^v

Finish realistic jobs in a live Linux shell: compile, configure, debug, and administer systems end to end.

Why it matters. This is what command-line coding agents actually do all day. It rewards models that recover from their own mistakes.

1. Claude Sonnet 5.570.6%^v
2. Claude Mythos 5.160.9%^v
3. Claude Opus 5.559.6%
4. GPT-6 Astra59.1%
5. Claude Fable 5.157.9%

The previous Terminal-Bench task set, kept for models that were never scored on 4.0.

Why it matters. Lets older and smaller models be compared on the same shell-agent tasks.

Watch out. Not comparable with Terminal-Bench 4.0 scores.

1. Claude Fable 5.191.4%
2. DeepSeek V4.1 Flash90.6%^v
3. MiMo-V2.6-Pro89.9%^v
4. Fugu Max89.5%^v
5. Claude Opus 589.1%

Solve competitive programming problems published after the model’s training cutoff.

Why it matters. Fresh problems limit memorization, so it tracks algorithmic skill honestly.

Watch out. Contest puzzles, not production code. Versions roll over, so check the window.

1. Fugu92.9%^v
2. Qwen3.8-Flash-Next91.9%^v
3. Qwen3.8-27B90.3%^v
4. Nemotron 3 Ultra 550B A55B89%^v
5. gpt-oss-120b87.8%

Write code that implements research-grade scientific computations across physics, chemistry, and biology.

Why it matters. Tests whether a model can turn domain knowledge into working numerical code.

1. Claude Opus 5.566.9%
2. Claude Fable 5.163.1%
3. Muse Wonder Full63%
4. Gemini 4 Argon62%
5. Claude Fable 561%

Act as a support agent: follow policy, call tools, and guide a simulated user through retail, airline, and telecom tasks.

Why it matters. Closest public proxy for customer-facing agents that must follow rules under pressure.

Watch out. Vendors often report a single domain. We note which one.

1. GLM-5.299.1%
2. Claude Fable 598.5%
3. Step 3.7 Flash98.5%
4. GLM-598.2%
5. Grok 4.397.7%

Find hard-to-locate facts on the open web through persistent multi-step browsing.

Why it matters. The standard for deep-research agents. It rewards persistence and search strategy.

Watch out. Heavily dependent on the search tool and time budget each vendor gives the model.

1. GPT-6 Astra91.5%^v
2. Kimi K391.2%^v
3. GPT-5.6 Sol90.4%^v
4. Gemini 3.1 Pro Preview85.9%^v
5. Claude Sonnet 584.7%^v

Operate a real desktop through screenshots, mouse, and keyboard to finish office and web tasks.

Why it matters. The main measure of computer-use agents that drive apps the way people do.

1. Kimi K384.8%^v
2. Qwen3.8-27B84.3%^v
3. MiMo-V2.6-Pro82%^v
4. Claude Opus 5.581.8%^v
5. Claude Sonnet 581.2%^v

Produce real professional deliverables (memos, spreadsheets, slide decks) across 44 occupations, graded head to head.

Why it matters. The most direct evidence of whether a model can do paid knowledge work.

1. Muse Wonder Full1857
2. Claude Opus 5.51846
3. Claude Sonnet 5.51844
4. Claude Fable 5.11735
5. Claude Opus 51708

Complete multi-step tasks that span many real MCP tool servers.

Why it matters. Tests tool discovery and orchestration across the MCP ecosystem most agents now use.

1. Kimi K384.2%^v
2. Inkling-Small79.6%^v
3. Beam78.7%^v
4. GLM-5.276.8%^v
5. Kimi K2.7 Code76%^v

Answer about 2,500 expert-written questions across more than a hundred subjects, chosen because earlier frontier models failed them.

Why it matters. The hardest broad knowledge-and-reasoning test with room left at the top.

Watch out. Scores with web search or code tools run much higher. We note which one was used.

1. Claude Sonnet 5.564.5%^v
2. Claude Opus 5.561.4%
3. Claude Fable 5.159.1%
4. Gemini 4 Argon57%
5. Fugu Ultra v256.3%^v

Answer graduate-level biology, chemistry, and physics questions that skilled non-experts get wrong even with web access.

Why it matters. A quick read on scientific reasoning. Useful for separating small and mid-size models.

Watch out. Frontier models now cluster near the ceiling, so small gaps at the top mean little.

1. GPT-6 Astra96.1%
2. Fugu Ultra v295.5%^v
3. Fugu Max95.5%^v
4. Fugu95.5%^v
5. Gemini 3.8 Flash95.3%

Answer 12,000 ten-option questions across 14 academic and professional fields.

Why it matters. A broad knowledge check that still separates small open-weight models.

Watch out. Saturated at the frontier.

1. Nemotron 3 Ultra 550B A55B86.8%^v
2. Gemma 4 31B85.2%^v
3. MAI-Thinking-185%^v
4. Nemotron 3 Super 120B A12B83.9%^v
5. Gemma 4 26B A4B82.6%^v

Infer the rule behind novel grid puzzles from a few examples, at a capped cost per task.

Why it matters. It targets fluid reasoning on problems the model cannot have seen. Humans solve these easily and models still struggle.

Watch out. Only ARC Prize–verified runs on the semi-private set are comparable.

1. GPT-6 Astra95%
2. Claude Opus 5.593.3%
3. GPT-5.6 Sol92.5%
4. Claude Opus 590.4%
5. Claude Fable 5.190%

Learn and beat unfamiliar interactive games with no instructions.

Why it matters. Measures skill acquisition in new environments rather than stored knowledge.

1. GPT-6 Astra62.7%
2. Claude Opus 530.2%
3. Gemini 3.8 Flash10.4%
4. GPT-5.6 Sol7.8%
5. Grok 4.62.1%

Solve the 2026 American Invitational Mathematics Examination problems.

Why it matters. Competition math is a sensitive test of multi-step reasoning in small models.

Watch out. Only 30 problems, so one question moves the score by more than 3 points.

1. ERNIE 5.199.6%^v
2. GLM-5.299.2%^v
3. Beam97.8%^v
4. Inkling97.1%^v
5. Kolibri-196%^v

Solve problems from the February 2026 Harvard-MIT Mathematics Tournament.

Why it matters. Harder than AIME, so it still separates reasoning models.

Watch out. Small problem set; treat single-point gaps as noise.

1. GLM-5.292.5%^v
2. Inkling-Small90.2%^v
3. Granite 4.2 30B89.2%^v
4. Inkling86.3%^v
5. MAI-Thinking-184.9%^v

Solve unpublished, research-level physics problems written by working physicists.

Why it matters. A test of genuine research reasoning. Most models still score in single digits.

1. Muse Wonder Full33%
2. GPT-5.6 Sol32.3%
3. GPT-6 Astra31.7%
4. Claude Opus 5.531.7%
5. GPT-6 Sol30.9%

Answer college-level questions that require reading charts, diagrams, and images.

Why it matters. The standard check for models that need to read screenshots, figures, and scans.

1. Claude Opus 5.587.7%
2. GPT-6 Astra86.9%
3. Gemini 3.8 Flash85.6%
4. Gemini 3.7 Flash85.5%
5. Claude Opus 584.7%

The original multi-discipline image-and-text reasoning set.

Why it matters. Still reported for small and open vision models.

Watch out. Easier than MMMU-Pro; do not mix the two.

1. Command A+75.1%^v

Answer questions about short, medium, and hour-long videos.

Why it matters. The main evidence for video understanding, not generation.

1. Kimi K390%^v
2. Doubao Seed 2.1 Pro89.2%^v
3. Doubao Seed 2.1 Turbo89%^v
4. Doubao Seed 2.0 Lite89%^v

Reason across sets of long documents, around 100,000 tokens each, to answer one question.

Why it matters. A 1M-token window is only useful if the model can use it. This checks that.

1. Kimi K388.7%
2. Step 5 Preview88.3%
3. MiMo-V2.6-Pro86.3%
4. Claude Fable 5.185.3%
5. Claude Opus 5.584.7%

Blind head-to-head votes from real users on which answer they prefer.

Why it matters. Captures tone, helpfulness, and formatting, which benchmarks miss.

Watch out. Rewards style as well as substance, and vendors test private variants before launch.

1. Gemini 4 Argon1525
2. Claude Fable 51506
3. Muse Spark 1.21500
4. Claude Fable 5.11499
5. Muse Spark 1.31493

Users vote between two generated web apps built from the same prompt.

Why it matters. A taste test for front-end and UI generation.

1. Claude Opus 5.51818
2. GPT-6 Astra1792
3. GPT-6.1 Sol1755
4. Claude Fable 5.11755
5. Claude Opus 51692

Blind votes between two clips with sound, generated from the same text prompt.

Why it matters. Most new video models generate dialogue, effects, and music with the picture. This board judges the whole clip.

Watch out. Elo is relative to the models on this board, so it cannot be compared with the silent-video board. Scores are from the v2.0 board, which uses a new scale anchored to Kling 3.0 Pro at 1000; older v1 results appear only on model pages.

1. Wan 3.01156
2. Utopai X1149
3. Dreamina Seedance 2.51143
4. MiniMax H31137
5. MiniMax H3 Max1130

Blind votes between two silent clips generated from the same text prompt.

Why it matters. Judges motion, composition, and prompt-following without sound to sway the vote.

Watch out. Elo is relative to the models on this board, so it cannot be compared with the with-audio board. Scores are from the v2.0 board, anchored to Kling 3.0 Pro at 1000; older v1 results appear only on model pages.

1. Wan 3.01129
2. MiniMax H3 Max1118
3. MiniMax H31117
4. Dreamina Seedance 2.51105
5. FLUX 3 Video1091

Blind votes between two clips with sound, animated from the same starting image.

Why it matters. Product shots and brand work usually start from an image, not a prompt.

1. MiniMax H3 Max1194
2. MiniMax H31181
3. Gemini Omni 1.1 Flash1178
4. Vidu Q4 Preview1178
5. Dreamina Seedance 2.01175

Blind votes between two silent clips animated from the same starting image.

Why it matters. Tests how faithfully a model keeps the subject of the source image while it moves.

1. Gemini Omni 1.1 Flash1369
2. Wan 3.01362
3. VBach 1.0 Pro1362
4. MiniMax H31357
5. PixVerse V61340

Blind votes between two images generated from the same prompt.

Why it matters. An independent ranking of image generators by what people actually prefer.

1. GPT Image 2.5 Sunburst1196
2. GPT Image 2.5 Flare1190
3. GPT Image 21171
4. Nano Banana 2.11160
5. Grok Imagine Image 2.01155

Blind votes on how well a model applies a requested edit to an existing image.

Why it matters. Editing is now most of real-world image work.

1. GPT Image 2.5 Sunburst1181
2. GPT Image 2.5 Flare1162
3. Nano Banana 2.11140
4. MAI-Image-2.61135
5. MAI-Image-2.6-Flash1125