llmrun Composite Score · Reasoning

The best LLMs for reasoning, open and proprietary together, on one open-ended scale where 100 is the average model (reasoning). Switch to open models to see only what you can download and run.

40 benchmarks4 sourcesUpdated 3 Oct 2026Reference scale 2026-Q4How the Composite Score is built

Reasoning score vs VRAM needed

Proprietary models are not shown on this chart.

Open models Overall

General Coding Agents Math Science Reasoning

Models ranked by reasoning llmrun Composite Score
#	Model	llmrun Composite Score	Coding	Agents	Math	Science	Reasoning	Benchmarks	VRAM
1	GPT 6 Astra Max19 benchmarks		156	144	152	142	163	19	–
2	GPT 6.1 Sol Max13 benchmarks		127	–	153	141	157	13	–
3	Claude Opus 5.5 Max17 benchmarks		147	–	152	144	154	17	–
4	Claude Sonnet 5.5 Max13 benchmarks		141	–	146	143	151	13	–
5	GPT 5.6 Sol Promax6 benchmarks		–	–	–	–	148	6	–
6	GPT 5.6 Sol Max21 benchmarks		146	141	143	142	145	21	–
7	Claude Fable 5.1 Max15 benchmarks		149	–	149	143	143	15	–
8	GPT 6 Sol Max18 benchmarks		149	–	147	141	143	18	–
9	Claude Opus 5 Max22 benchmarks		146	144	143	139	143	22	–
10	GPT 5.4 Pro (Mar 05, 2026)8 benchmarks		–	–	136	141	142	8	–
11	GPT 6 Astra6 benchmarks		–	164	–	–	142	6	–
12	Claude Opus 5.54 benchmarks		–	–	–	145	141	4	–
13	Claude Fable 5 Max20 benchmarks		147	134	148	138	141	20	–
14	Claude Fable 5.19 benchmarks		147	–	–	142	141	9	–
15	Claude Opus 59 benchmarks		147	–	–	139	141	9	–
16	Claude Fable 56 benchmarks		–	–	–	–	140	6	–
17	GPT 5.59 benchmarks		141	–	–	137	139	9	–
18	GPT 5.6 Sol9 benchmarks		142	–	–	140	139	9	–
19	GPT 5.6 Terra Max20 benchmarks		139	138	140	139	138	20	–
20	Gemini 3 Deep Think Preview3 benchmarks	—	–	–	–	–	137	3	–
21	GPT 5.5 Pro7 benchmarks		–	–	142	–	137	7	–
22	Gemini 3.7 Flash15 benchmarks		118	–	130	129	137	15	–
23	Gemini 3.1 Pro Preview8 benchmarks		–	–	–	–	135	8	–
24	GPT 6 Sol4 benchmarks		–	–	–	139	134	4	–
25	GPT 5.4 (Mar 05, 2026)20 benchmarks		134	129	133	134	134	20	–
26	Claude Opus 4.8 Max19 benchmarks		127	127	136	129	133	19	–
27	Gemini 3.5 Flash4 benchmarks		–	–	–	–	132	4	–
28	GPT 5.6 Terra10 benchmarks		135	–	–	136	132	10	–
29	Gemini 3.8 Flash6 benchmarks		133	–	–	125	132	6	–
30	DeepSeek V4 Pro 0813open991 GB · 21 benchmarks		125	–	128	126	131	21	991 GB
31	Claude Opus 4.7 Max15 benchmarks		131	–	130	120	130	15	–
32	Gemini 3.6 Flash16 benchmarks		113	–	124	125	130	16	–
33	Kimi K3open1670 GB · 26 benchmarks		139	133	135	136	129	26	1670 GB
34	Grok 4.79 benchmarks		–	–	122	130	129	9	–
35	Grok 4.2011 benchmarks		116	122	–	–	129	11	–
36	DeepSeek V4 Flash 0731open183 GB · 21 benchmarks		121	–	128	124	129	21	183 GB
37	Qwen3.8 Max10 benchmarks		–	–	132	–	129	10	–
38	Muse Spark 1.27 benchmarks		121	–	–	131	128	7	–
39	GPT 5.6 Luna Max20 benchmarks		134	135	135	130	128	20	–
40	GPT 5.2 (Dec 11, 2025)12 benchmarks		–	–	124	–	128	12	–
41	Muse Spark 1.3 Max6 benchmarks		–	–	132	138	128	6	–
42	Grok 4.516 benchmarks		128	–	125	128	127	16	–
43	Claude Sonnet 4.68 benchmarks		–	116	–	–	127	8	–
44	GPT 6 Luna Max18 benchmarks		134	–	135	128	127	18	–
45	Claude Sonnet 4.6 Max11 benchmarks		–	–	128	108	126	11	–
46	GLM 5.3open454 GB · 21 benchmarks		131	140	129	131	126	21	454 GB
47	Qwen3.7 Max17 benchmarks		121	–	126	122	125	17	–
48	GPT 5.6 Luna8 benchmarks		–	–	–	129	125	8	–
49	Gemini 3 Flash Preview7 benchmarks		–	–	–	–	125	7	–
50	Claude Sonnet 5 Max13 benchmarks		–	–	132	114	125	13	–
51	DeepSeek V4.1 Flashopen458 GB · 10 benchmarks		116	–	131	124	125	10	458 GB
52	Claude Opus 4.85 benchmarks		–	136	–	–	124	5	–
53	Claude Opus 4.5 (Nov 01, 2025)9 benchmarks		–	125	–	–	123	9	–
54	GPT 6 Luna5 benchmarks		131	–	–	127	123	5	–
55	Claude Opus 4.6 Max12 benchmarks		–	–	128	124	123	12	–
56	Claude Opus 4.77 benchmarks		132	147	–	–	122	7	–
57	Inkling Smallopen163 GB · 12 benchmarks		–	–	117	118	122	12	163 GB
58	Gemini 3 Pro Preview17 benchmarks		124	127	120	126	122	17	–
59	Inklingopen575 GB · 21 benchmarks		104	–	112	116	121	21	575 GB
60	Muse Spark 1.36 benchmarks		–	–	131	–	121	6	–
61	Grok 4.311 benchmarks		–	–	117	118	120	11	–
62	Qwen3.6 Max Preview8 benchmarks		–	–	–	–	120	8	–
63	GLM 5.2open454 GB · 24 benchmarks		120	132	126	128	120	24	454 GB
64	Kimi K2.6open618 GB · 18 benchmarks		119	–	122	122	119	18	618 GB
65	GPT 5 Pro (Oct 06, 2025)5 benchmarks		–	–	123	–	119	5	–
66	Grok 4.1 Fast Reasoning6 benchmarks		–	–	–	–	119	6	–
67	Qwen3.8 27Bopen17.9 GB · 9 benchmarks		112	–	118	112	119	9	17.9 GB
68	GPT 5.1 (Nov 13, 2025)14 benchmarks		121	–	–	–	119	14	–
69	Kimi K2.7 Codeopen618 GB · 19 benchmarks		115	119	121	118	118	19	618 GB
70	Grok 4.20 0309 Reasoning8 benchmarks		–	–	118	–	118	8	–
71	DeepSeek V4 Proopen960 GB · 22 benchmarks		111	–	118	122	117	22	960 GB
72	Grok 4 (Jul 09)12 benchmarks		111	103	–	–	117	12	–
73	GPT 5 (Aug 07, 2025)20 benchmarks		120	–	122	118	117	20	–
74	GPT 5.4 Mini (Mar 17, 2026)12 benchmarks		–	–	120	118	116	12	–
75	Claude Sonnet 4.5 (Sep 29, 2025)11 benchmarks		–	115	–	105	115	11	–
76	NVIDIA Nemotron 3 Ultra 550B A55B BF16open378 GB · 14 benchmarks		103	–	112	112	115	14	378 GB
77	Qwen3.7 Plus9 benchmarks		103	–	–	117	115	9	–
78	GLM 5.1open454 GB · 13 benchmarks		114	127	117	117	115	13	454 GB
79	O3 (Apr 16, 2025)19 benchmarks		114	–	112	110	115	19	–
80	Claude Opus 4 (May 14, 2025)7 benchmarks		–	–	–	96	115	7	–
81	Claude Opus 4.1 (Aug 05, 2025)6 benchmarks		–	–	–	–	115	6	–
82	Qwen3.5 397B A17Bopen274 GB · 9 benchmarks		–	–	112	–	115	9	274 GB
83	Qwen3.7 Flash5 benchmarks		–	–	109	–	115	5	–
84	Kimi K2.5open618 GB · 20 benchmarks		110	111	–	117	114	20	618 GB
85	MiniMax M2.5open140 GB · 11 benchmarks		109	110	–	–	114	11	140 GB
86	DeepSeek V3.2 Expopen413 GB · 10 benchmarks		107	–	–	101	113	10	413 GB
87	O3 Pro (Jun 10, 2025)6 benchmarks		117	–	–	–	113	6	–
88	Qwen3.5 122B A10Bopen88.1 GB · 6 benchmarks		–	–	–	96	112	6	88.1 GB
89	Gemini 3.5 Flash Lite11 benchmarks		103	–	106	–	112	11	–
90	DeepSeek V3.2open413 GB · 13 benchmarks		95	109	–	–	112	13	413 GB
91	Qwen3.5 Flash10 benchmarks		–	–	106	–	112	10	–
92	Gemini 2.5 Pro13 benchmarks		–	102	109	112	112	13	–
93	DeepSeek V4 Flashopen175 GB · 10 benchmarks		107	–	–	111	112	10	175 GB
94	Qwen3.5 Plus11 benchmarks		–	–	–	–	112	11	–
95	GPT 5 Mini (Aug 07, 2025)15 benchmarks		111	–	117	–	111	15	–
96	MiMo V2.5 Proopen616 GB · 6 benchmarks		116	–	–	115	111	6	616 GB
97	GPT 5.4 Nano (Mar 17, 2026)6 benchmarks		–	–	–	115	111	6	–
98	O4 Mini (Apr 16, 2025)19 benchmarks		110	–	112	108	111	19	–
99	Qwen3.5 27Bopen17.9 GB · 6 benchmarks		98	–	–	–	111	6	17.9 GB
100	Claude 3.7 Sonnet (Feb 19, 2025)4 benchmarks		–	–	–	–	110	4	–
101	Grok 4 Fast6 benchmarks		–	–	–	–	110	6	–
102	Qwen3.6 Plus15 benchmarks		108	–	116	114	110	15	–
103	Gemma 4 31B ITopen21.7 GB · 11 benchmarks		108	–	–	105	109	11	21.7 GB
104	GLM 5open454 GB · 15 benchmarks		110	118	–	–	109	15	454 GB
105	Claude Haiku 4.5 (Oct 01, 2025)8 benchmarks		–	–	104	–	109	8	–
106	DeepSeek V3.1 Terminusopen412 GB · 5 benchmarks		108	–	–	103	108	5	412 GB
107	Tiny Recursion Model2 benchmarks	—	–	–	–	–	107	2	–
108	MiniMax M3open258 GB · 19 benchmarks		113	113	106	118	107	19	258 GB
109	Gemini 3.1 Flash Lite6 benchmarks		–	–	109	–	107	6	–
110	Qwen3.6 35B A3Bopen22.0 GB · 12 benchmarks		105	–	109	110	107	12	22.0 GB
111	Qwen3.6 27Bopen17.9 GB · 12 benchmarks		102	–	114	111	106	12	17.9 GB
112	Gemini 2.5 Pro Preview (Mar 25)6 benchmarks		–	–	–	–	106	6	–
113	GLM 5.3 Flashopen205 GB · 17 benchmarks		113	135	121	124	106	17	205 GB
114	Qwen Plus (Apr 28, 2025)2 benchmarks	—	–	–	–	–	106	2	–
115	Qwen3 235B A22B Thinking 2507open143 GB · 13 benchmarks		103	–	–	108	106	13	143 GB
116	Qwen3.5 35B A3Bopen22.0 GB · 8 benchmarks		–	–	–	108	106	8	22.0 GB
117	DeepSeek V3.1open412 GB · 7 benchmarks		–	–	–	–	105	7	412 GB
118	Qwen3 Max (Sep 23, 2025)13 benchmarks		–	–	105	–	105	13	–
119	Gemini 2.5 Flash Preview (May 20)6 benchmarks		–	–	–	–	105	6	–
120	GLM 4.7open218 GB · 13 benchmarks		101	106	111	111	104	13	218 GB

The range after each score is a 90% interval: given the boards a model has results on, its true score is very likely inside it. Short ranges mean many agreeing results; long ranges mean few or conflicting ones.

Why some models are missing. A model gets a score only with results on at least 4 benchmarks across at least 2 categories, one of them reasoning or coding. Skill scores need at least 2 benchmarks in that skill; otherwise they show as –.

VRAM is llmrun's estimate at Q4_K_M (or the smallest quantization we track) for the weights plus room for a 16K-token context. Longer contexts need more memory. Proprietary models can't run locally, so the VRAM filter hides them.

How the llmrun Composite Score is built · llmrun does not run these benchmarks; scores are aggregated from public sources.