AI Model Release Tracker
Coding

SWE-Bench Verified

Real coding tasks pulled from open-source projects — the AI has to find and fix actual bugs. A human-checked version of the original SWE-Bench. Higher is better.

Rankings

Higher is better
#1
Claude Opus 5
via BenchLM
96%
2
Claude Fable 5
95.5%
2
Claude Mythos 5
via BenchLM
95.5%
4
Claude Opus 4.8
via BenchLM
88.6%
5
Claude Opus 4.7
via BenchLM
87.6%
6
Claude Sonnet 5
via BenchLM
85.2%
7
GPT-5.3-Codex
via BenchLM
85%
8
Claude Opus 4.5
via BenchLM
80.9%
9
Claude Opus 4.6
80.8%
10
Gemini 3.1 Pro
80.6%
10
DeepSeek-V4-Pro-0813OW
via BenchLM
80.6%
12
Qwen3.7-Max
via BenchLM
80.4%
13
Kimi K2.6
80.2%
14
GPT-5.2
80%
15
Claude Sonnet 4.6
79.6%
16
DeepSeek-V4-Flash-0731OW
via BenchLM
79%
17
Qwen3.6-Plus
via BenchLM
78.8%
18
Gemini 3.0 Flash
78%
19
GLM-5
77.8%
20
Qwen3.7-Plus
via BenchLM
77.7%
21
Mistral Medium 3.5
77.6%
22
Muse Spark
77.4%
23
Claude Sonnet 4.5
77.2%
24
Kimi K2.5
76.8%
25
Grok 4.20 Beta
via BenchLM
76.7%
26
Qwen3.5OW
76.4%
27
GPT-5.1
76.3%
28
Gemini 3.0 Pro
76.2%
29
Muse GlimmerOW
76%
30
Claude Opus 4.1
74.5%
31
GLM-4.7
73.8%
32
Qwen3.6OW
73.4%
33
Claude Haiku 4.5
73.3%
34
Claude Sonnet 4
72.7%
35
Claude Opus 4
72.5%
36
Devstral 2 (123B)
72.2%
37
Nemotron 3 UltraOW
via BenchLM
71.9%
38
Kimi K2 Thinking
71.3%
39
Nemotron 3.5 Super VLOW
71.1%
40
Grok Code Fast 1
via BenchLM
70.8%
41
Qwen3-Coder-NextOW
70.6%
42
DeepSeek V3.1 Terminus
68.4%
43
GLM-4.6
68%
43
Devstral Small 2 (24B)
68%
45
DeepSeek V3.1
via BenchLM
66%
46
Kimi K2
65.8%
46
Kimi K2 (0905)
65.8%
48
GLM-4.5
64.2%
49
gpt-oss-120bOW
62.4%
50
Claude 3.7 Sonnet
62.3%
51
gpt-oss-20bOW
60.7%
52
Nemotron 3 SuperOW
60.5%
53
Kimi-Dev
60.4%
54
GLM-4.5-Air
59.8%
55
Gemini 2.5 Pro
59.6%
56
DeepSeek R1-0528
57.6%
57
GPT-4.1
via BenchLM
54.6%
58
Nemotron 3.5 LightningOW
51.6%
59
o3-mini
via BenchLM
49.3%
60
Claude 3.5 Sonnet (upgraded)
49%
61
Devstral SmallOW
46.8%
62
DeepSeek V3 Chat
42%
63
Claude 3.5 Haiku
40.6%
64
Claude 3.5 Sonnet
33.4%
65
Claude 3 Opus
33%
66
Gemini 2.5 Flash-Lite
via BenchLM
31.6%
67
GPT-4.1 mini
via BenchLM
23.6%
68
Gemini 2.0 Flash
via BenchLM
21.4%
69
Gemini 2.0 Flash-Lite
via BenchLM
12.5%
70
Gemini 1.5 Flash
via BenchLM
9.6%

Scores marked “via” above are quoted with attribution from BenchLM, retrieved 19 October 2026.

SWE-Bench Verified — frequently asked questions

Real coding tasks pulled from open-source projects — the AI has to find and fix actual bugs. A human-checked version of the original SWE-Bench. Higher is better.