TPSTokens Per Second

AI Model Benchmarks

389 scores across 12 benchmarks — quality, coding, reasoning, and speed.

Output throughput in tokens per second measured on standard prompts. Higher = faster responses. Source: artificialanalysis.ai, ~Feb 2026.

Model
tok/s
1
Gemini 3.1 Flash-LiteGoogle DeepMind
360
2
Step-3.5-Flash-BaseStepFun
300
3
Gemini 2.0 FlashGoogle DeepMind
228
4
GPT-5.4 NanoOpenAI
210
5
Gemini 2.5 FlashGoogle DeepMind
198
6
Claude Haiku 4.5Anthropic
187
7
SonarPerplexity AI
175
8
Gemma 4 26B A4B ITGoogle DeepMind
163.09
9
Llama 4 ScoutMeta AI
160
10
Grok 4 minixAI
145
11
Grok 3 minixAI
138
12
GPT-5.4 miniOpenAI
132
13
Sonar ReasoningPerplexity AI
120
14
o4-miniOpenAI
115
15
GPT-4.1OpenAI
112
16
o3-miniOpenAI
108
17
GPT-4oOpenAI
98
18
Sonar ProPerplexity AI
95
19
Qwen3.5-35B-A3BAlibaba / Qwen
92.51
20
Llama 4 MaverickMeta AI
90
21
GPT-5.4OpenAI
85
22
Claude Sonnet 4.5Anthropic
82
23
Claude Sonnet 4.6Anthropic
80
24
Claude 3.5 SonnetAnthropic
78
25
Mistral Medium 3Mistral AI
75
26
Mistral Large 2Mistral AI
72
27
DeepSeek V3DeepSeek
70
28
Claude 3.7 SonnetAnthropic
68
29
Qwen3 72BAlibaba / Qwen
65
30
Sonar Reasoning ProPerplexity AI
65
31
Gemini 2.5 ProGoogle DeepMind
62
32
Phi-4Microsoft
60
33
DeepSeek R1DeepSeek
58
34
Grok 3xAI
55
35
QwQ-32BAlibaba / Qwen
50
36
Gemini 1.5 ProGoogle DeepMind
45
37
GPT-4.5OpenAI
42
38
Claude Opus 4Anthropic
40
39
Claude Opus 4.6Anthropic
38
40
o3OpenAI
32
41
Grok 4xAI
28
42
Llama 3.1 405BMeta AI
26
43
Grok 4.20xAI
24
44
o1OpenAI
22
45
Sonar Deep ResearchPerplexity AI
18

Quality scores from official model cards and published technical reports. Speed benchmarks from artificialanalysis.ai (approximate medians). Higher is better for all metrics.