Model Rankings

Model	Overall	Reasoning	Coding	Agentic coding	Mathematics	Data analysis	Language	Instruction following	Elo
1Claude Fable 5max-effort	83.0	89.7	86.0	62.2	96.0	80.5	90.7	75.8	1506#1
2GPT-5.6 Solmax	81.1	91.7	83.9	56.2	96.2	79.8	87.7	71.8	n/a
3GPT-5.5xhigh	80.2	89.7	82.1	54.0	95.9	81.6	87.4	70.7	1477#20
4Claude Opus 5max-effort	80.1	91.2	81.4	65.2	95.7	74.6	88.7	63.8	n/a
5Kimi K3	79.2	90.7	81.4	62.2	84.4	78.7	85.5	71.4	n/a
6GPT-5.4xhigh	78.0	88.1	77.5	53.8	94.1	79.3	82.6	70.2	1465#39
7Muse Spark 1.2xhigh	78.0	90.0	77.5	57.6	91.2	76.5	78.6	74.3	n/a
8GPT-5.6 Terramax	77.9	90.6	78.2	54.9	94.9	79.3	82.9	64.6	n/a
9Grok 4.6xhigh	77.6	91.2	75.0	54.2	93.1	75.3	82.7	71.5	n/a
10Gemini 3.1 Propreview-high	77.0	84.0	76.5	44.1	91.0	78.5	85.4	79.1	1486#13
11Claude Opus 4.7xhigh-effort	76.5	87.2	82.1	50.7	92.9	78.3	77.9	66.7	1494#7
12Claude Opus 4.8max-effort	76.2	89.2	81.8	50.5	94.3	66.0	79.7	72.0	1473#27
13Claude Sonnet 5xhigh-effort	76.0	88.7	80.7	59.4	92.9	71.7	75.0	63.9	n/a
14Grok 4.5	75.8	87.2	68.6	56.5	90.8	73.0	82.8	71.5	1469#35
15Muse Spark 1.1xhigh	75.3	87.7	77.2	58.5	87.1	72.5	74.3	69.6	1489#10
16Gemini 3.5 Flashhigh	74.6	82.0	78.2	49.0	88.2	64.9	84.6	75.6	n/a
17GPT-5.2high	74.6	83.2	76.1	50.3	93.2	78.2	79.8	61.8	1435#83
18Claude Opus 4.6thinking-auto-high-effort	74.5	88.7	78.2	49.0	89.3	69.9	83.3	63.3	1498#5
19DeepSeek V4 Flash 0731	74.2	86.6	75.0	46.8	86.8	79.3	79.2	65.5	n/a
20GPT-5.2 Codex	74.0	77.7	83.6	49.4	88.8	78.2	73.7	66.4	n/a
21Gemini 3.6 Flashhigh	73.6	85.1	77.9	43.4	86.4	63.0	83.9	75.4	1484#15
22GPT-5.6 Lunamax	73.6	85.6	82.9	48.4	87.2	78.0	72.6	60.1	n/a
23GLM-5.2	73.2	78.6	79.7	51.8	89.8	73.7	76.2	62.3	n/a
24Qwen3.7-Max	73.1	83.3	74.2	43.6	85.2	71.8	79.7	74.0	1474#25
25Claude Sonnet 4.6thinking-auto-medium-effort	73.0	84.8	79.3	42.6	87.0	77.9	76.1	63.2	1472#30
26Claude Opus 4.5thinking-64k-high-effort	72.6	80.1	79.7	39.7	90.4	74.4	81.3	62.5	n/a
27Inklingxhigh	71.9	78.3	71.0	49.4	88.4	72.8	73.5	70.1	1442#73
28DeepSeek-V4-Pro	71.6	82.7	70.0	42.6	90.7	74.5	78.1	62.4	1458#50
29Kimi K2.6thinking	70.5	79.4	78.6	46.9	84.3	65.1	75.1	64.4	1461#44
30GPT-5.4 Nanoxhigh	69.6	81.1	70.8	46.8	91.0	67.6	62.5	67.2	n/a

Two sources, never merged. Everything left of Elo is LiveBench, scored 0–100 and machine-marked; Elo is LMArena human preference. They share neither method nor scale, so agreement is real corroboration rather than one measurement counted twice. Shading marks the top five per column. The effort suffix matters — a max-effort run is a different measurement from the same model at low effort, so it is shown rather than hidden. LiveBench CC-BY-SA-4.0; LMArena CC-BY-4.0.

21 of 120 rated models led the field on their release date. The vertical spread at any date is how much the field varies at one moment; the upward drift is progress.

A single Elo number hides this. Two models a point apart overall can sit twenty percentiles apart on coding or on Japanese, which is the difference that matters when you are choosing one.

• Qwen3-0.6B27,629,624
• Qwen3-8B15,207,936
• Qwen3.5-9B12,423,514
• RoBERTa Base12,221,106
• Qwen2.5-1.5B12,108,126
• RoBERTa Large11,249,715
• Qwen2.5-7B11,218,582
• Gemma 4 26B A4B9,985,859
• Gemma 4 31B IT9,882,137
• Llama 3.2 1B9,198,251

A closed model has no repository, so GPT and Claude cannot appear here at all. This ranks the most-downloaded OPEN models, not the most-used models — there is no admissible free source for proprietary popularity, and inventing a proxy for it would be a confident number measuring nothing. Counts include automated pulls.