artificialwatch

✳ the bench index

The LLM leaderboard: every benchmark, one table.

The public boards each tell part of the story — community votes here, live task suites there, lab-reported launch numbers somewhere else. This table reads them together: each model's best-variant score on every board we track, normalized to a percentile rank, and combined into one composite index. Refreshed daily, free, no signup. Two models side by side? See the head-to-head comparisons. For cost, there's every model's API price in one table, and for coding alone, the AI coding leaderboard, and for models you can download, the open-weights leaderboard. Just the answer? The best AI model right now.

Table last refreshed 2026-10-28 · each board's own as-of date is listed with the sources below.

The index at a glance

trailingleading
Claude Opus 5.5100
GPT-6 Astra92
Claude Fable 5.190.9
Claude Sonnet 5.589.8
GPT-6.1 Sol88.7
Muse Wonder82.9
Gemini 4 Argon80.9
GPT-6.1 Terra75
Kimi K3.1 Pro73.9
Muse Spark 1.373
Claude Opus 572.8
Claude Fable 571.4
GPT-6 Sol68.6
GPT-5.566.7
DeepSeek V4.1 Flash60.8
Grok 4.657.4
Qwen 3.8 Max56.5
Grok 4.755.7
GPT-5.652.3
Claude Haiku 5.550
Gemini 3.1 Pro48.6
GLM-5.348
Gemini 3.8 Flash44
MiMo-V2.6 Pro43.5
Claude Opus 4.836.4
MiniMax M3.1 Flash36
Grok 4.534.3
Muse Spark 1.130.5
DeepSeek V4 Pro30.4
Mistral Large 428
Claude Sonnet 527.3
GPT-6 Luna26.9
Kimi K326.1
GLM-5.225.7
Muse Spark Open24
Qwen 3.7 Max22.9
Gemini 3.5 Flash21.1
GLM-5.3 Flash17.4
Kimi K2.610.8
Beam8.7
Kimi K2.7 Code6.8
MiniMax M35.7
Inkling4.3
#ModelAW IndexLMArenaLiveBenchSimpleBenchLLM StatsVals AIAA
1Claude Opus 5.5Anthropic1001,808.7 10083.2 9788.4 10060.4 10070.2 9658.4 100
2GPT-6 AstraOpenAI921,773.5 9882.2 9186.5 9358.2 8365 8356.9 96
3Claude Fable 5.1Anthropic90.91,745 9183.4 10086.6 9653.2 38—49.7 72
4Claude Sonnet 5.5Anthropic89.81,769.3 9677.8 5775.9 5758.4 9270.9 10054.2 88
5GPT-6.1 SolOpenAI88.71,751.6 9381.6 8982.9 8953.6 5863.9 7854.5 92
6Muse WonderMeta82.91,726.1 8981.2 83—53.2 3866.8 8749.7 76
7Gemini 4 ArgonGoogle80.91,679.9 82——55.1 7569.2 9150.4 80
8GPT-6.1 TerraOpenAI751,670.7 75———60.2 6551 84
9Kimi K3.1 ProMoonshot73.91,694.5 86———62.5 7446 68
10Muse Spark 1.3Meta731,656.2 6881.6 8681.8 7852.2 17——
11Claude Opus 5Anthropic72.81,690.3 8480.1 7180.6 7453.8 67——
12Claude Fable 5Anthropic71.41,623.9 6183 9481.9 8253.1 25——
13GPT-6 SolOpenAI68.61,677.4 8079.2 6973.1 52———
14GPT-5.5OpenAI66.71,512.1 1880.2 7476.9 67———
15DeepSeek V4.1 FlashDeepSeek60.81,666.3 7381.1 8066.7 3751.2 061.3 7032.5 52
16Grok 4.6xAI57.41,616.7 5778 6075.9 57———
17Qwen 3.8 MaxAlibaba56.51,676 7778.5 6362.5 26—59 5733.6 56
18Grok 4.7xAI55.71,637.2 6677.4 51——58.6 4835 60
19GPT-5.6OpenAI52.31,615.8 5581.1 7771.7 4853.3 50——
20Claude Haiku 5.5Anthropic501,587.8 4872.1 20——58.7 5236.2 64
21Gemini 3.1 ProGoogle48.61,480.2 977 4979.6 70———
22GLM-5.3Zhipu AI481,622.7 5976.1 4366.2 33—59.7 6131.7 48
23Gemini 3.8 FlashGoogle441,583.6 4675.8 3782.4 85—56 3930.5 44
24MiMo-V2.6 ProXiaomi43.51,625.5 64———57.8 4429.9 40
25Claude Opus 4.8Anthropic36.41,555.8 3676.2 4664.8 30———
26MiniMax M3.1 FlashMiniMax361,581.1 41———53.8 3528.9 36
27Grok 4.5xAI34.31,553.1 3475.8 3470 41———
28Muse Spark 1.1Meta30.51,542.4 3075.3 31————
29DeepSeek V4 ProDeepSeek30.41,581.9 4377.4 5450.9 7—53 3023.1 20
30Mistral Large 4Mistral281,542.7 3271.8 11———25.3 28
31Claude Sonnet 5Anthropic27.31,540.3 2776 4060.6 19———
32GPT-6 LunaOpenAI26.91,579.8 3972 17——52.3 2227.8 32
33Kimi K3Moonshot26.11,656.2 7179.2 6660.7 2251.9 853 26—
34GLM-5.2Zhipu AI25.71,605.6 5073.2 2658.8 15———
35Muse Spark OpenMeta241,525.7 25———48.6 1324.7 24
36Qwen 3.7 MaxAlibaba22.91,515.1 2173.1 2370.4 44———
37Gemini 3.5 FlashGoogle21.11,499.1 1474.6 2976.7 63——11.7 0
38GLM-5.3 FlashZhipu AI17.41,606.7 5271.6 9——50.9 17—
39Kimi K2.6Moonshot10.81,509.3 1670.5 6————
40BeamReflection AI8.71,436.1 2———43.4 917.7 12
41Kimi K2.7 CodeMoonshot6.81,473.7 768.4 357.9 11———
42MiniMax M3MiniMax5.71,483.3 1167.3 045.8 0——18.7 16
43InklingThinking Machines4.31,442 571.9 1450 4—29 412.5 4
44Inkling SmallThinking Machines01,413.4 0———27.6 012.9 8
—Seed 2.1 ProByteDanceneeds 2+ boards1,519.1 23—————

Cell format: raw board score + percentile chip. Hover any cell for the exact variant behind the number.

How the index is computed

01Best variant wins

Labs ship many configs — thinking modes, effort levels, harnesses. Per board, each model is scored by its best-performing variant, identically for everyone. Search, grounding, image and video product modes are excluded.

02Percentiles, not raw averages

Elo ≈1500 and 0–100 scores can't be averaged. Each board's scores become percentile ranks (0–100) within the set of models on this table that the board actually scores.

03The index is a median

A model's AW Index is the median of its percentiles across boards — so no single board's methodology can dominate. A model needs 2+ boards to be ranked; boards scoring fewer than 5 of these models display but don't vote.

04Nothing is estimated

Every raw number comes from the source board, links back to it, and carries its fetch date. Percentiles are relative to this table's model set, not the boards' full lists.

The boards

LMArena ↗community votes — millions of blind head-to-headsElo (community votes) · 45 models mapped · as of 2026-10-28LiveBench ↗independent — contamination-free live task suiteglobal average (0-100) · 36 models mapped · as of 2026-10-18SimpleBench ↗independent — private reasoning set, human-baselinedAVG@5 % (private set) · 28 models mapped · as of 2026-10-28LLM Stats ↗LLM Stats' headline composite — verified benchmarks + live perfLLM Stats Score · composite (0-100), their rendered top 15 · 13 models mapped · as of 2026-10-28Vals AI ↗independent — expert-crafted private tasks, measured runsVals Index accuracy % · weighted finance + coding tasks · 24 models mapped · as of 2026-10-28Artificial Analysis ↗independent — standardized eval suite, measured runs onlyintelligence index · measured runs only · 26 models mapped · as of 2026-10-28

The data

This table is free to use: JSON · CSV, rebuilt every day. The index and the percentile ranks are ours, under CC BY 4.0; each board's raw scores belong to that board, and the JSON names and links every source. All our open data.

Cite asArtificialWatch Bench Index, 2026-10-28. ArtificialWatch. https://artificialwatch.com/benchmarks

Quick answers

What is the best AI model right now?

By the public leaderboards combined, Claude Opus 5.5: it ranks first on the ArtificialWatch Bench Index at 100, ahead of GPT-6 Astra (92) and Claude Fable 5.1 (90.9). The index takes the median of each model's percentile across the boards that score it (it reads 6), so no single board decides it; each board's own number is in the table.

Which leaderboards does this ranking combine?

LMArena, LiveBench, SimpleBench, LLM Stats, Vals AI and Artificial Analysis. Each model's best-variant score on each board is converted to a percentile within that board, and a model needs at least 2 boards to get an index score.

How often is the ranking updated?

Every day: the boards are re-read each morning and the table is rebuilt. This version was composed October 28, 2026.

Benchmarks tell you who leads today; the radar tells you the minute that changes. Watch the models · the wire · new releases · head-to-heads · coding leaderboard · open weights · API pricing · track record

This page and the launch alerts carry one sponsor at a time. Sponsor ArtificialWatch