{ }AICoder
English 中文
June 2026

AI Coder Bench

Real-world coding benchmark rankings. 50 tasks across bug fixing, feature building, refactoring, system design, and debug & explain.

7Boards
54Models ranked
777Tasks
August 2026Period
2026-08-13Last update
●We fetch raw rankings from 6 public benchmark providers directly — no re-weighting, no composite score. Click "Open official board" on each card to go to the source.

📌 7 official boards · mirrored 1:1

Each card previews a provider's Top 5. Click "View full" for the complete board, or jump to the official page. Data is 100% from each provider — we do no aggregation or math.

LMArena Agent ★
Human preference on real agentic coding sessions
Agent
Metric:Agent Score· 27 models
  1. #1Claude Fable 512.6
  2. #2Grok 4.610.5
  3. #3GPT-5.6 Sol10.11
  4. #4Kimi K39.75
  5. #5Claude Opus 4.89.5
Official ↗
CursorBench
Cursor's first-party agentic coding benchmark (v3.2)
Agent
Metric:Pass rate%· 17 models
  1. #1Claude Fable 572.9%
  2. #2Grok 4.669.9%
  3. #3GPT-5.6 Sol67.2%
  4. #4Grok 4.566.7%
  5. #5GPT-5.6 Terra64.9%
Official ↗
Artificial Analysis
Composite Intelligence Index over 20+ benchmarks
Overall
Metric:AA Index· 49 models
  1. #1Grok 4.661
  2. #2Claude Opus 560.7
  3. #3Claude Fable 559.9
  4. #4GPT-5.6 Sol58.9
  5. #5Kimi K357.1
Official ↗
LiveBench
Contamination-resistant benchmark, refreshed monthly
Reasoning
Metric:Avg score%· 27 models
  1. #1GPT-5.6 Sol82.9%
  2. #2Grok 4.682.1%
  3. #3Claude Fable 581.4%
  4. #4Claude Opus 580.7%
  5. #5GPT-5.580.5%
Official ↗
Aider Polyglot
133 real coding tasks across 6 languages
Coding
Metric:Pass rate%· 20 models
  1. #1GPT-5.6 Sol88.0%
  2. #2Grok 4.686.5%
  3. #3GPT-5.584.9%
  4. #4Gemini 2.5 Pro83.1%
  5. #5Grok 4.579.6%
Official ↗
LiveCodeBench
Continuously updated competitive-programming set
Reasoning
Metric:Pass@1%· 12 models
  1. #1DeepSeek-V4-Pro100.0%
  2. #2GPT-5.4100.0%
  3. #3GPT-5.4 mini100.0%
  4. #4GPT-5.5100.0%
  5. #5Claude Sonnet 4.6100.0%
Official ↗
EvalPlus
HumanEval+ strict test suite
Coding
Metric:HumanEval+%· 18 models
  1. #1GPT-5.489.0%
  2. #2GPT-5.4 mini89.0%
  3. #3Grok 4.688.4%
  4. #4Qwen3.6-Plus87.2%
  5. #5DeepSeek-V4-Pro86.6%
Official ↗

LMArena AgentAgent

Metric: Agent Score·27 models·fetched 2026-08-13

Open official board ↗
#ModelAgent ScoreCross-board ranks
#1
Claude Fable 5Anthropic
12.6
C #1
A #3
L #3
A—L—E—
#2
Grok 4.6xAI
10.5
C #2
A #1
L #2
A #2
L #12
E #3
#3
GPT-5.6 SolOpenAI
10.11
C #3
A #4
L #1
A #1
L—
E #18
#4
Kimi K3Moonshot AI
9.75
C—
A #5
L #8
A—L—E—
#5
Claude Opus 4.8Anthropic
9.5
C #6
A #6
L #7
A #7
L #7
E #10
#6
Claude Sonnet 5Anthropic
8.61
C #8
A #10
L #12
A—L—E—
#7
GPT-5.5OpenAI
8.61
C #11
A #8
L #5
A #3
L #4
E—
#8
GLM-5.2Zhipu AI
7.12
C #13
A #13
L #18
A—L—E—
#9
Grok 4.5xAI
5.47
C #4
A #9
L #11
A #5
L—
E #8
#10
GPT-5.4OpenAI
5.38
C—
A #11
L #9
A #9
L #2
E #1
#11
GPT-5.6 TerraOpenAI
3.96
C #5
A #7
L #6
A—L—E—
#12
GPT-5.6 LunaOpenAI
3.28
C #9
A #12
L #15
A—L—E—
#13
Claude Sonnet 4.6Anthropic
3.14
C #15
A #16
L #17
A #8
L #5
E #7
#14
GLM-5.1Zhipu AI
0.72
C—
A #24
L—A—L—E—
#15
Kimi K2.7 CodeMoonshot AI
0.38
C #14
A #22
L #22
A—L—E—
#16
Qwen3.7-MaxQwen
-0.17
C—
A #18
L #16
A #10
L #10
E—
#17
Gemini 3.1 ProGoogle
-0.31
C—
A #17
L #10
A—L—E—
#18
Gemini 3 ProGoogle
-0.61
C—
A #15
L #14
A—L—E—
#19
DeepSeek-V4-ProDeepSeek
-0.72
C #7
A #20
L #19
A #6
L #1
E #5
#20
Gemini 3.5 FlashGoogle
-0.94
C #16
A #14
L #13
A—L—E—
#21
Kimi K2.6Moonshot AI
-1
C #17
A #21
L #20
A #11
L—E—
#22
MiniMax-M3MiniMax
-2.55
C—
A #19
L #24
A—L—E—
#23
DeepSeek-V4-FlashDeepSeek
-2.95
C #10
A #23
L #26
A—L—
E #6
#24
Grok 4.3xAI
-7.87
C—
A #29
L #27
A #13
L #11
E—
#25
Grok Build 0.1xAI
-7.97
C—
A #27
L #23
A—L—E—
#26
MiniMax-M2.7MiniMax
-11.11
C—
A #28
L—A—L—E—
#27
Nemotron-UltraNvidia
-13.08
C—A—L—A—L—E—
Rank color:#1#2#3#4-5#6-10#11+

The "Cross-board ranks" column shows this model's rank on the other boards at a glance (grey = not tested there). Boards use different methodologies, so raw values are not directly comparable.

LMArena Agent board: real agentic coding sessions scored by human blind votes (signal-based agent score). Sourced from arena.ai/leaderboard/agent and mirrored 1:1 — we recompute nothing.

📊 7 benchmark providers
LMArena Agent ↗ Agent

LMArena Agent board: real agentic coding sessions scored by human blind votes (signal-based agent score). Sourced from arena.ai/leaderboard/agent and mirrored 1:1 — we recompute nothing.

Agent Score27 models
CursorBench ↗ Agent

CursorBench is Cursor's first-party agentic coding benchmark (v3.2). Tasks come from real Cursor sessions with ambiguous, multi-file requirements. Metric: pass rate (%). Mirrored from Cursor's official cursor.com/cursorbench snapshot.

Pass rate%17 models
Artificial Analysis ↗ Overall

Artificial Analysis aggregates 20+ benchmarks (MMLU-Pro, GPQA, HLE, SciCode, IFBench, Terminal-bench, etc.) into a single Intelligence Index — the most comprehensive commercial evaluator for "overall capability".

AA Index49 models
LiveBench ↗ Reasoning

LiveBench refreshes its questions monthly to prevent contamination. Covers reasoning, math, coding, language, instruction following, and data analysis.

Avg score%27 models
Aider Polyglot ↗ Coding

133 real coding tasks (bug-fix / feature) across 6 languages. Metric: pass_rate_2.

Pass rate%20 models
LiveCodeBench ↗ Reasoning

Continuously updated competitive-programming problems. Metric: Pass@1.

Pass@1%12 models
EvalPlus ↗ Coding

HumanEval+ strict test suite, 164 tasks.

HumanEval+%18 models

All rankings and numbers come directly from official sources; we only do name matching and aggregate presentation, no weighting or re-ranking. Boards use different methodologies, so do not compare raw values across boards.