Skip to content
Model Pareto

Hard reasoning models ranked by quality, price and speed

BenchmarkScienceMathOther

Quality vs price

13 models on the frontier — nothing else is both cheaper and better.

Lower price is better. Pareto frontier: Gemma 4 E4B, Sarvam 30B, Qwen3.5 4B, Sarvam 105B, Ling 3.0 Flash, MiMo-V2.5, GPT-6 Luna, Qwen3.8 Flash-Next, GLM-5.3 Flash, MiMo-V2.6-Pro, Muse Spark 1.3, Claude Sonnet 5.5, Claude Opus 5.5.

  • Best-value frontier (nothing is both cheaper and better)
  • Evidencestrong → weak
  • Estimated from other categories
  • Estimated (price or speed)

Ranking · 182 models

Price $/M tokens (blended 3:1) · Speed output tokens/s

Hard reasoning ranking. Column headers sort the table. Expand a row to see per-benchmark scores.
Select to compareSourceDetails
1
Claude Opus 5.5Frontier
99.9±2.6$8.00922/4 benchmarks measuredVerified
2
Claude Fable 5.1
95.7±3.7$20.0663/4 benchmarks measuredVerified
3
Claude Sonnet 5.5Frontier
95.3±4.9$4.00est.~792/4 benchmarks measuredVerified
4
GPT-6 Astra
93.5±3.7$20.0523/4 benchmarks measuredVerified
5
Claude Opus 5
91.7±3.7$10.0543/4 benchmarks measuredVerified
6
Muse Spark 1.3Frontier
86.9±3.7$2.002193/4 benchmarks measuredVerified
7
MiMo-V2.6-ProFrontierOpen
86.1±5.1$0.54492/4 benchmarks measuredVerified
8
GPT-6 Sol
85.8±5.1$4.001102/4 benchmarks measuredVerified
9
Kimi K3Open
83.4±3.7$6.00373/4 benchmarks measuredVerified
10
Step 5 Preview
83.2±4.0$1.43853/4 benchmarks measuredMixed
11
Gemini 3.8 Flash
82.9±3.7$1.502913/4 benchmarks measuredVerified
12
Grok 4.6
82.2±3.7$3.00603/4 benchmarks measuredVerified
13
Qwen3.8 Max (0902)
82.1±3.7$3.00393/4 benchmarks measuredVerified
14
Grok 4.7
81.8±5.1$3.00402/4 benchmarks measuredVerified
15
GLM-5.3Open
81.0±3.7$2.15573/4 benchmarks measuredVerified
16
GPT-5.6 Terra
80.1±3.7$4.50833/4 benchmarks measuredVerified
17
GPT-5.5
79.9±3.7$11.3913/4 benchmarks measuredVerified
18
Hy4 previewOpen
78.9±6.6$1.25est.~782/4 benchmarks measuredLab-reported
19
Qwen3.8 2.4T A95BOpen
78.8±3.7$3.00383/4 benchmarks measuredVerified
20
GPT-5.3 Codex
78.5±5.8$4.811452/4 benchmarks measuredVerified
21
GLM-5.3 FlashFrontierOpen
77.9±3.7$0.24463/4 benchmarks measuredVerified
22
Claude Sonnet 5
76.5±3.7$4.00793/4 benchmarks measuredVerified
23
DeepSeek V4.1 FlashOpen
76.1±4.0$0.522323/4 benchmarks measuredMixed
24
Qwen3.8 Flash-NextFrontierOpen
76.0±3.7$0.23543/4 benchmarks measuredVerified
25
Gemini 3.1 Pro
75.8±3.7$4.501203/4 benchmarks measuredVerified
26
DeepSeek V4 Pro (0813)Open
75.6±3.7$1.98973/4 benchmarks measuredVerified
27
Qwen3.8-Omni-Flash
73.5±6.6$0.23est.~542/4 benchmarks measuredLab-reported
28
GPT-6 LunaFrontier
72.4±5.1$0.201322/4 benchmarks measuredVerified
29
Grok Build 0.1 0616
72.0±5.8$1.25552/4 benchmarks measuredVerified
30
MiniMax-M3Open
70.7±3.7$0.521183/4 benchmarks measuredVerified
31
Qwen3.8 27BOpen
69.8±3.7$1.13433/4 benchmarks measuredVerified
32
Motif 3Open
69.3±5.8self-host~$0.13–1.19est.~962/4 benchmarks measuredVerified
33
Inkling SmallOpen
68.6±5.8$0.522062/4 benchmarks measuredVerified
34
Nex-N2-ProOpen
68.3±5.8self-host~$0.20–1.76est.~922/4 benchmarks measuredVerified
35
K2 Horizon 375B A23BOpen
65.8±3.7self-host~$0.17–1.38est.~883/4 benchmarks measuredVerified
36
Qwen3.7 Plus
65.6±3.7$0.70653/4 benchmarks measuredVerified
37
Kimi K2.7 CodeOpen
65.4±3.7$1.71663/4 benchmarks measuredVerified
38
Solar Pro 4
64.7±5.8$0.52932/4 benchmarks measuredVerified
39
InklingOpen
64.6±3.6$1.761844/4 benchmarks measuredMixed
40
Apodex 1.1
64.3±3.7$0.97est.~1003/4 benchmarks measuredVerified
41
Hy3Open
64.1±3.7$0.25863/4 benchmarks measuredVerified
42
A.X-K2Open
63.9±5.8self-host~$0.25–2.21est.~842/4 benchmarks measuredVerified
43
Solar Open2 250BOpen
63.4±5.8self-host~$0.12–1.11est.~942/4 benchmarks measuredVerified
44
MiMo-V2.5FrontierOpen
61.2±5.8$0.18402/4 benchmarks measuredVerified
45
GPT-5.4 mini
60.2±3.7$1.692173/4 benchmarks measuredVerified
46
Xing4.0-29B-A4BOpen
59.8±10.0self-host~$0.067–0.60est.~1101/4 benchmarks measuredLab-reported
47
Gemma 4 31BOpen
59.6±4.5$0.21353/4 benchmarks measuredMixed
48
Nemotron 3 UltraOpen
58.9±3.7$1.072153/4 benchmarks measuredVerified
49
Ling 3.0 FlashFrontierOpen
58.7±5.8$0.113462/4 benchmarks measuredVerified
50
Qwen3.5 397B A17BOpen
57.7±3.7$1.35863/4 benchmarks measuredVerified
51
Ling-3.0-flash-VLOpen
56.5±3.7$0.111473/4 benchmarks measuredVerified
52
LongCat 2.0Open
56.2±3.7$0.52est.~793/4 benchmarks measuredVerified
53
K2 Horizon MoVA 36B A4BOpen
56.1±3.7self-host~$0.053–0.44est.~1103/4 benchmarks measuredVerified
54
GPT-5.4 nano
55.9±3.7$0.461653/4 benchmarks measuredVerified
55
o3
55.4±5.8$3.501432/4 benchmarks measuredVerified
56
Step 3.7 FlashOpen
55.4±5.8$0.441902/4 benchmarks measuredVerified
57
GPT-5.5 Instant (June 2026)
54.3±3.7$11.31233/4 benchmarks measuredVerified
58
Muse GlimmerOpen
54.0±3.6$0.581334/4 benchmarks measuredMixed
59
Intern-S2-Preview (35B-A3B)Open
53.6±8.6self-host~$0.046–0.38est.~1201/4 benchmarks measuredLab-reported
60
Qwen3.5 122B A10BOpen
52.5±3.7$1.101293/4 benchmarks measuredVerified
61
K-EXAONE 2.0 0803Open
52.3±5.8self-host~$0.29–2.58est.~822/4 benchmarks measuredVerified
62
Qwen3.6 35B A3BOpen
52.0±3.7$0.841203/4 benchmarks measuredVerified
63
Gemini 3.5 Flash-Lite
51.7±3.7$0.853513/4 benchmarks measuredVerified
64
Gemma 4 26B A4BOpen
51.6±5.8$0.16est.~892/4 benchmarks measuredVerified
65
KAT Coder Pro V2
51.6±5.8$0.52est.~1102/4 benchmarks measuredVerified
66
Ling-3.0-flash-FinOpen
51.3±5.1$0.111592/4 benchmarks measuredVerified
67
Ring-2.6-1TOpen
51.1±3.7$0.851193/4 benchmarks measuredVerified
68
Quasar 438B
50.2±3.7$0.901523/4 benchmarks measuredVerified
69
Nova 2.0 Lite
50.1±4.3$0.851673/4 benchmarks measuredVerified
70
G9v3-39A5BOpen
50.1±5.8self-host~$0.061–0.50est.~1102/4 benchmarks measuredVerified
71
Qwen3.5 Omni Plus
50.0±5.8$1.50822/4 benchmarks measuredVerified
72
Qwen3.5 9BOpen
48.2±5.8$0.15602/4 benchmarks measuredVerified
73
gpt-oss-120bOpen
47.8±3.4$0.261864/4 benchmarks measuredVerified
74
K2 Horizon 7BOpen
47.1±3.7self-host~$0.083–0.74est.~1003/4 benchmarks measuredVerified
75
Phi-4-reasoning-plusOpen
46.9±7.5self-host~$0.12–1.08est.~482/4 benchmarks measuredLab-reported
76
Nova 2.0 Pro Preview
46.7±4.3$3.441203/4 benchmarks measuredVerified
77
Nemotron 3 Super 120B A12BOpen
45.6±3.7$0.451643/4 benchmarks measuredVerified
78
EXAONE 4.5 33BOpen
45.5±5.8self-host~$0.17–1.49est.~842/4 benchmarks measuredVerified
79
Nanbeige4.1-3BOpen
45.2±5.8self-host~$0.047–0.42est.~1102/4 benchmarks measuredVerified
80
HyperNova 60B 2605Open
44.5±5.8self-host~$0.066–0.59est.~1102/4 benchmarks measuredVerified
81
Mistral Medium 3.5Open
43.9±3.6$3.001494/4 benchmarks measuredMixed
82
Nemotron Cascade 2 30B A3BOpen
42.0±5.8self-host~$0.044–0.37est.~2102/4 benchmarks measuredVerified
83
Cogito v2.1Open
41.7±5.8$1.25est.~822/4 benchmarks measuredVerified
84
INTELLECT-3Open
41.6±5.8self-host~$0.12–0.97est.~972/4 benchmarks measuredVerified
85
Claude Haiku 4.5
41.0±3.4$2.001094/4 benchmarks measuredVerified
86
Mistral Small 4Open
40.5±3.6$0.261704/4 benchmarks measuredMixed
87
Apriel-v1.6-15B-ThinkerOpen
39.7±5.8self-host~$0.12–1.09est.~942/4 benchmarks measuredVerified
88
Sarvam 105BFrontierOpen
39.0±5.8$0.07est.~992/4 benchmarks measuredVerified
89
Trinity Large ThinkingOpen
38.9±3.7$0.393373/4 benchmarks measuredVerified
90
Ling-2.6-1TOpen
38.9±5.8$0.85est.~872/4 benchmarks measuredVerified
91
gpt-oss-20bOpen
38.6±3.4$0.101814/4 benchmarks measuredVerified
92
Command A+Open
38.4±3.7self-host~$0.19–1.562143/4 benchmarks measuredVerified
93
K2 Horizon 3.7BOpen
38.3±3.7self-host~$0.056–0.50est.~1103/4 benchmarks measuredVerified
94
Nova 2.0 Omni
38.2±5.8$0.85est.~1702/4 benchmarks measuredVerified
95
Hermes 4 - Llama-3.1 405BOpen
37.7±5.8$1.50422/4 benchmarks measuredVerified
96
Ling 3.0 TinyOpen
36.7±5.8self-host~$0.029–0.24312/4 benchmarks measuredVerified
97
K2 Think V2Open
36.4±5.8self-host~$0.36–3.22est.~742/4 benchmarks measuredVerified
98
Step3 VL 10BOpen
36.4±5.8self-host~$0.091–0.81est.~1202/4 benchmarks measuredVerified
99
Magistral Small 1.2Open
36.3±4.3$0.75est.~1103/4 benchmarks measuredVerified
100
Nemotron 3.5 LightningOpen
36.2±3.7$0.112673/4 benchmarks measuredVerified
101
Qwen3 Next 80B A3B InstructOpen
35.6±5.8$0.411582/4 benchmarks measuredVerified
102
DiffusionGemma 26B A4BOpen
35.5±5.8self-host~$0.062–0.56est.~892/4 benchmarks measuredVerified
103
Qwen3.5 4BFrontierOpen
35.5±5.8$0.06232/4 benchmarks measuredVerified
104
Ring-flash-2.0Open
35.3±5.8$0.25est.~1202/4 benchmarks measuredVerified
105
K2-V2Open
35.3±5.8self-host~$0.35–3.09est.~752/4 benchmarks measuredVerified
106
North Mini CodeOpen
35.2±3.7self-host~$0.044–0.37933/4 benchmarks measuredVerified
107
ERNIE 4.5 300B A47BOpen
34.9±5.8$0.49est.~792/4 benchmarks measuredVerified
108
NVIDIA Nemotron 3 Nano 30B A3BOpen
34.6±3.7$0.092453/4 benchmarks measuredVerified
109
Falcon-H1R-7BOpen
34.5±5.8self-host~$0.074–0.66est.~1002/4 benchmarks measuredVerified
110
Solar Open 100BOpen
34.1±5.8self-host~$0.12–0.95est.~972/4 benchmarks measuredVerified
111
Llama 3.1 Nemotron Ultra 253B v1Open
33.7±5.8self-host~$0.88–8.40est.~1102/4 benchmarks measuredVerified
112
Qwen3 Coder NextOpen
33.3±3.7$0.561263/4 benchmarks measuredVerified
113
MiniCPM5-2BOpen
33.2±3.7self-host~$0.035–0.31est.~1203/4 benchmarks measuredVerified
114
Granite 4.2 30BOpen
32.9±5.8$0.28772/4 benchmarks measuredVerified
115
Mercury 2.5
31.3±5.1$0.347812/4 benchmarks measuredVerified
116
Gemma 4 12BOpen
30.3±5.8$0.151182/4 benchmarks measuredVerified
117
Granite 4.2 8BOpen
29.8±3.7$0.11833/4 benchmarks measuredVerified
118
Sarvam 30BFrontierOpen
28.7±5.8$0.05est.~1202/4 benchmarks measuredVerified
119
LongCat Flash LiteOpen
28.5±5.8self-host~$0.059–0.52est.~1202/4 benchmarks measuredVerified
120
HyperCLOVA X SEED Think (32B)Open
26.9±5.8self-host~$0.17–1.50est.~832/4 benchmarks measuredVerified
121
Tri-21B-ThinkOpen
26.7±5.8self-host~$0.16–1.46est.~902/4 benchmarks measuredVerified
122
Mistral Large 3Open
25.4±3.4$0.75774/4 benchmarks measuredVerified
123
EXAONE 4.0 32BOpen
25.4±5.8self-host~$0.16–1.46est.~842/4 benchmarks measuredVerified
124
Qwen3 Omni 30B A3B InstructOpen
25.4±5.8$0.43962/4 benchmarks measuredVerified
125
Celeris-1
24.5±3.7$0.3315213/4 benchmarks measuredVerified
126
Olmo 3.1 32B ThinkOpen
24.1±5.8self-host~$0.16–1.46est.~842/4 benchmarks measuredVerified
127
NVIDIA Nemotron Nano 12B v2 VLOpen
23.4±5.8$0.302012/4 benchmarks measuredVerified
128
Granite 4.2 3BOpen
23.4±3.7$0.052203/4 benchmarks measuredVerified
129
Llama 4 MaverickOpen
22.9±4.3$0.421143/4 benchmarks measuredVerified
130
LFM2.5-2.6BOpen
22.4±5.8self-host~$0.037–0.33est.~1202/4 benchmarks measuredVerified
131
Devstral 2Open
22.1±3.4self-host~$0.53–5.07est.~844/4 benchmarks measuredVerified
132
Gemma 4 E4BFrontierOpen
21.6±5.8$0.04612/4 benchmarks measuredVerified
133
GLM-4.6VOpen
21.4±5.8$0.45682/4 benchmarks measuredVerified
134
LFM2.5-8B-A1BOpen
21.1±5.8self-host~$0.031–0.26est.~1302/4 benchmarks measuredVerified
135
NVIDIA Nemotron Nano 9B V2Open
20.9±5.8$0.091552/4 benchmarks measuredVerified
136
Phi-4Open
20.2±5.8$0.22402/4 benchmarks measuredVerified
137
Olmo 3.1 32B InstructOpen
20.2±5.8self-host~$0.16–1.46est.~842/4 benchmarks measuredVerified
138
Olmo 3 7B ThinkOpen
19.7±5.8self-host~$0.072–0.64est.~1002/4 benchmarks measuredVerified
139
NVIDIA Nemotron 3 Nano 4BOpen
19.4±5.8self-host~$0.047–0.42est.~2002/4 benchmarks measuredVerified
140
Ministral 3 14BOpen
19.4±3.4$0.20874/4 benchmarks measuredVerified
141
Devstral Small 2Open
19.3±3.4self-host~$0.19–1.67est.~1104/4 benchmarks measuredVerified
142
Reka Flash 3Open
18.6±5.8$0.35est.~892/4 benchmarks measuredVerified
143
Command AOpen
18.6±5.8$4.38572/4 benchmarks measuredVerified
144
Nemotron 3 Nano Omni 30B A3B ReasoningOpen
18.1±5.8$0.422752/4 benchmarks measuredVerified
145
Llama 4 ScoutOpen
17.4±4.3$0.311073/4 benchmarks measuredVerified
146
Phi-4-mini-flash-reasoningOpen
17.0±8.3self-host~$0.046–0.41est.~582/4 benchmarks measuredLab-reported
147
Falcon-H1-34B-InstructOpen
16.7±7.5self-host~$0.17–1.52est.~832/4 benchmarks measuredLab-reported
148
Qwen3.5 2BOpen
16.6±5.8self-host~$0.030–0.27est.~952/4 benchmarks measuredVerified
149
Ministral 3 8BOpen
16.5±3.4$0.15874/4 benchmarks measuredVerified
150
Hermes 4 - Llama-3.1 70BOpen
16.1±5.8self-host~$0.35–3.12est.~672/4 benchmarks measuredVerified
151
LFM2 24B A2BOpen
16.1±5.8self-host~$0.051–0.42est.~1202/4 benchmarks measuredVerified
152
Exaone 4.0 1.2BOpen
14.8±5.8self-host~$0.024–0.20est.~1402/4 benchmarks measuredVerified
153
Llama 3.2 Instruct 90B (Vision)Open
14.2±5.8self-host~$0.47–4.23est.~632/4 benchmarks measuredVerified
154
Olmo 3 7B InstructOpen
13.8±5.8$0.13est.~1002/4 benchmarks measuredVerified
155
Gemma 4 E2BOpen
13.6±5.8self-host~$0.035–0.32est.~962/4 benchmarks measuredVerified
156
Molmo2-8BOpen
13.4±5.8self-host~$0.081–0.72est.~1002/4 benchmarks measuredVerified
157
Granite 4.0 H SmallOpen
12.6±5.8$0.11482/4 benchmarks measuredVerified
158
Llama 3.3 Instruct 70BOpen
12.5±4.3$0.71853/4 benchmarks measuredVerified
159
LFM2.5-1.2B-ThinkingOpen
11.9±5.8self-host~$0.023–0.19est.~1402/4 benchmarks measuredVerified
160
LFM2.5-1.2B-InstructOpen
11.8±5.8self-host~$0.023–0.19est.~1402/4 benchmarks measuredVerified
161
Ministral 3 3BOpen
11.8±3.4$0.102314/4 benchmarks measuredVerified
162
MiniCPM5-1BOpen
11.7±5.8self-host~$0.021–0.18est.~1402/4 benchmarks measuredVerified
163
Jamba 1.7 LargeOpen
11.3±5.8self-host~$0.55–4.92est.~712/4 benchmarks measuredVerified
164
MiniCPM-V 4.6 1.3BOpen
11.1±5.8self-host~$0.024–0.20est.~1402/4 benchmarks measuredVerified
165
Kimi Linear 48B A3B InstructOpen
10.7±5.8self-host~$0.051–0.42est.~962/4 benchmarks measuredVerified
166
Nova Micro
10.5±5.8$0.062772/4 benchmarks measuredVerified
167
Granite 4.0 MicroOpen
10.0±5.8self-host~$0.043–0.38est.~1102/4 benchmarks measuredVerified
168
Jamba Reasoning 3BOpen
9.5±5.8self-host~$0.041–0.37est.~1202/4 benchmarks measuredVerified
169
Jamba 1.7 MiniOpen
9.1±5.8self-host~$0.10–0.90est.~972/4 benchmarks measuredVerified
170
Phi-4 Multimodal InstructOpen
9.1±5.8self-host~$0.060–0.53172/4 benchmarks measuredVerified
171
Tiny Aya GlobalOpen
9.0±5.8self-host~$0.042–0.38est.~1302/4 benchmarks measuredVerified
172
LFM2.5-VL-1.6BOpen
8.8±5.8self-host~$0.028–0.23est.~1302/4 benchmarks measuredVerified
173
Granite 4.0 H 350MOpen
8.8±5.8self-host~$0.011–0.088est.~1502/4 benchmarks measuredVerified
174
Apertus 70B InstructOpen
8.5±5.8$1.34est.~752/4 benchmarks measuredVerified
175
K2 Horizon 0.9BOpen
8.2±3.7self-host~$0.021–0.18est.~1403/4 benchmarks measuredVerified
176
Qwen3.5 0.8BOpen
8.0±5.8self-host~$0.018–0.15est.~1102/4 benchmarks measuredVerified
177
Granite 4.0 350MOpen
7.9±5.8self-host~$0.011–0.090est.~1502/4 benchmarks measuredVerified
178
Granite 4.0 H 1BOpen
7.9±5.8self-host~$0.026–0.21est.~1202/4 benchmarks measuredVerified
179
Apertus 8B InstructOpen
7.4±5.8$0.13est.~1002/4 benchmarks measuredVerified
180
Llama 3.2 Instruct 11B (Vision)Open
7.2±5.8$0.34152/4 benchmarks measuredVerified
181
Molmo 7B-DOpen
6.8±5.8self-host~$0.077–0.68est.~1002/4 benchmarks measuredVerified
182
Gemma 3 270MOpen
4.8±5.3self-host~$0.009–0.077est.~1302/4 benchmarks measuredVerified

About hard reasoning

Graduate-level science and competition math. Quality is a weighted average of the benchmarks below, each rescaled so 100 = the best tracked model. Benchmarks without a published score are estimated from related ones (shown dashed). Dot fill shows the evidence: solid = strong, tinted = medium, empty ring = weak (few benchmarks measured, lab-reported only or provisional). Methodology

Day-0 (model cards) Third-party

Top quality
Claude Opus 5.5
99.9 / 100
Cheapest on the frontier
Gemma 4 E4B
21.6 quality · $0.04
Newest
Claude Sonnet 5.5
Sep 28, 2026
Fully verified
92%
of 182 models