Benchmarks
347 sources · 1,258 models × 3,038 benchmarks · 158,912 scoresArtificial Analysis Coding Agent Index v2.1 scores coding-agent and model pairs, such as Claude Code or Codex with a given model and effort, as the equal-weight average of mean scores over three attempts per task on FrontierCode-AA 2.0 (132 tasks; each attempt scores 0 or between 70% and 100%), Terminal-Bench 4.1 (66 tasks) and SWE-Atlas-QnA-AA (124 tasks, grading adjusted by Artificial Analysis). FrontierCode-AA 2.0 replaces v2.0's FrontierCode v1.1 and Terminal-Bench 4.1 replaces Terminal-Bench 4.0, so v2.1 scores are not comparable with v2.0. Scores belong to the agent harness and model together. It is separate from the model-level Artificial Analysis Coding Index.
No signal or difficulty estimate. Too few models have been measured on this benchmark to estimate its difficulty.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high · Antigravity CLI | 55.5 | |
| 2 | Muse Wonder Full | max · Muse Code | 54.5 | |
| 3 | GPT-6 Astra | max · Codex | 53.1 | |
| 4 | GPT-6.1 Sol | max · Codex | 52.2 | |
| 5 | GPT-6.1 Terra | max · Codex | 48.6 | |
| 6 | Kimi K3.1 Pro | max · Kimi Code | 48.6 | |
| 7 | Gemini 4 Argon | high · Antigravity CLI | 45.5 | |
| 8 | Muse Wonder | max · Muse Code | 44.5 | |
| 9 | Grok 4.7 | x-high · Grok Build | 41.6 | |
| 10 | Mistral Large 4 | high · Mistral Vibe | 40.6 | |
| 11 | GLM-5.3 | max · OpenCode | 40.1 | |
| 12 | Muse Ember | max · Muse Code | 39.7 | |
| 13 | DeepSeek V4.1 Flash 1017 | max · DeepSeek Harness | 36.8 | |
| 14 | Muse Spark 1.3 | max · Muse Code | 36.8 | |
| 15 | Step 5 Preview | not reported · Claude Code | 35.8 | |
| 16 | GPT-6 Luna | max · Codex | 25.9 | |
| 17 | Beam 1.0 501B A23B | not reported · Mirror | 15.5 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).