SWE-Bench-Pro agent results
CodeAgentBench
A leaderboard for coding agents on SWE-Bench-Pro: 151 tasks, 3 attempts per task, and 453 planned scoreable attempts in total. Rankings include completed 453/453 runs and sort by Final Score by default; partial progress rows are shown at the bottom when available.
Best Final Score
37.59
OpenCode - GLM 5.2
Completed Models
32
No partial progress rows
Scoreable Attempts
453
151 tasks x 3 tries
Exported
2026-06-18
SWEPro 151 zh pass@3
Benchmark Leaderboard
Final Score is the primary ranking signal. It weights Pass^3 most strongly for stable agent performance, while Pass@3 keeps each model's best-of-three reach visible.
Agent
Notes:
Final Score = 100 × S^0.55 × R^0.25 × A^0.20, where S=(Pass^3)^(1/3), R=1-(1-Pass@3)^(1/3), and A=Attempt Score.
Pass^3 counts tasks solved in all 3 attempts.
Pass@3 counts tasks solved at least once across 3 attempts.
Attempt Score is solved scoreable attempts divided by 453.
Tokens are shown in K and summed across all exported attempts. Input includes cache read/write tokens when exported. OpenCode rows use per-step command.log usage when available; other rows fall back to per-task run_summary.json fields. -- means non-zero usage was not exported.
Agent Version combines each run's exported runtime label with the current local CLI version command when historical reports do not include exact semver.
Completed rows are rankable only when the exported summary reports 453 scoreable attempts; partial rows are shown as progress references at the bottom.
Visual Leaderboard
Switch metrics to compare the same agent/model pairs by reach, consistency, and per-attempt solve rate.
GLM 5.2
OpenCode - Zhipu GLM
GPT 5.5 (xhigh)
Codex - OpenAI
GPT 5.4 (xhigh)
Codex - OpenAI
GPT 5.3 codex (xhigh)
Codex - OpenAI
Kimi K2.6(Kimi for Coding)
Kimi - Moonshot
MiniMax M3
OpenCode - MiniMax
DeepSeek v4 pro (max)
OpenCode - DeepSeek
GLM 5.1
OpenCode - Zhipu GLM
Qwen 3.7 Max (1m)
Qoder - Qwen
Qwen 3.5 plus
Qwen - Qwen
Qwen 3.7 Plus (1m)
Qoder - Qwen
MiMo v2.5 pro
Claude Code - Xiaomi
MiMo v2.5 pro (high)
OpenCode - Xiaomi
MiniMax M2.5 highspeed
OpenCode - MiniMax
DeepSeek v4 flash (max)
OpenCode - DeepSeek
GLM 5 turbo
OpenCode - Zhipu GLM
Qwen 3.6 plus
Qwen - Qwen
Step 3.7 flash
OpenCode - StepFun
KAT Coder Pro v2
OpenCode - StreamLake
MiMo v2.5
Claude Code - Xiaomi
Qwen 3.5 plus (180k)
Qoder - Qwen
Qwen 3.6 plus (180k)
Qoder - Qwen
MiniMax M2.7 highspeed
OpenCode - MiniMax
MiMo v2.5
OpenCode - Xiaomi
LongCat 2.0 Preview
OpenCode - LongCat
Step 3.5 flash 2603
OpenCode - StepFun
GLM 4.7
OpenCode - Zhipu GLM
KAT Coder Pro v2
OpenCode - StreamLake
SenseNova 6.7 flash lite
OpenCode - SenseNova
doubao seed 2.0 code
OpenCode - Volcengine
Step 3.5 flash
OpenCode - StepFun
DeepSeek v4 flash (max)
deepseek-tui - DeepSeek
Reach vs Consistency
Pass@3 shows whether an agent can solve a task at least once; Pass^3 shows whether it solves the same task all three times. The gap is useful when comparing stochastic or retry-sensitive agents.
Pass@3
OpenCode - GLM 5.2
Codex - GPT 5.5 (xhigh)
Codex - GPT 5.4 (xhigh)
Codex - GPT 5.3 codex (xhigh)
Kimi - Kimi K2.6(Kimi for Coding)
OpenCode - MiniMax M3
OpenCode - DeepSeek v4 pro (max)
OpenCode - GLM 5.1
Qoder - Qwen 3.7 Max (1m)
Qwen - Qwen 3.5 plus
Qoder - Qwen 3.7 Plus (1m)
Claude Code - MiMo v2.5 pro
OpenCode - MiMo v2.5 pro (high)
OpenCode - MiniMax M2.5 highspeed
OpenCode - DeepSeek v4 flash (max)
OpenCode - GLM 5 turbo
Qwen - Qwen 3.6 plus
OpenCode - Step 3.7 flash
OpenCode - KAT Coder Pro v2
Claude Code - MiMo v2.5
Qoder - Qwen 3.5 plus (180k)
Qoder - Qwen 3.6 plus (180k)
OpenCode - MiniMax M2.7 highspeed
OpenCode - MiMo v2.5
OpenCode - LongCat 2.0 Preview
OpenCode - Step 3.5 flash 2603
OpenCode - GLM 4.7
OpenCode - KAT Coder Pro v2
OpenCode - SenseNova 6.7 flash lite
OpenCode - doubao seed 2.0 code
OpenCode - Step 3.5 flash
deepseek-tui - DeepSeek v4 flash (max)
Pass^3