Track frontier AI model rankings across SWE-bench, LMArena, Terminal-Bench, METR, coding, reasoning, and agent benchmarks in one crawlable leaderboard index.
Artificial Analysis composite coding score: equally-weighted average of SciCode, Terminal-Bench Hard, and LiveCodeBench. Higher = better.
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 59.1% |
| 2 | GPT-5.4 | 57.2% |
| 3 | Claude Opus 4 | 56.7% |
| 4 | Gemini 3.1 Pro | 55.5% |
| 5 | GPT-5.3 Codex | 53.1% |
| 6 | Claude Opus 4.7 | 53.1% |
| 7 | Claude Sonnet 4.6 | 50.9% |
| 8 | GPT-5.2 | 48.7% |
| 9 | Claude Opus 4.6 | 48.1% |
| 10 | Claude Opus 4.5 | 47.8% |
Crowdsourced human preference Elo ratings. All models.
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 1501 |
| 2 | Claude Opus 4.7 | 1488 |
| 3 | Gemini 3.1 Pro | 1482 |
| 4 | Gemini 3 Pro | 1480 |
| 5 | GPT-5.4 | 1471 |
| 6 | GLM-5 | 1470 |
| 7 | GPT-5 | 1468 |
| 8 | Gemini 3 Flash | 1466 |
| 9 | Claude Opus 4 | 1461 |
| 10 | Gemini 2.5 Pro | 1457 |
LMArena Elo filtered to non-proprietary licenses. Where OSS actually stands.
| # | Model | Score |
|---|---|---|
| 1 | GLM-5 | 1470 |
| 2 | Kimi K2 | 1455 |
| 3 | GLM-4.6 | 1441 |
| 4 | GLM-4.7 | 1436 |
| 5 | Mistral Large | 1430 |
| 6 | DeepSeek R1 | 1428 |
| 7 | DeepSeek V3 | 1424 |
| 8 | DeepSeek V3.1 | 1420 |
| 9 | Qwen 3 235B | 1419 |
| 10 | Qwen 3 Coder | 1356 |
Length of task (in minutes) an AI can complete at ~50% success, from METR's Time Horizon 1.1 suite (HCAST + SWAA). Longer = more capable.
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 11h59m |
| 2 | Gemini 3.1 Pro | 6h24m |
| 3 | GPT-5.2 | 5h52m |
| 4 | GPT-5.3 Codex | 5h50m |
| 5 | GPT-5.4 | 5h42m |
| 6 | Claude Opus 4.5 | 4h53m |
| 7 | Gemini 3 Pro | 3h44m |
| 8 | GPT-5.1 Codex Max | 3h44m |
| 9 | GPT-5 | 3h23m |
| 10 | o3 | 1h60m |
Where developers are actually spending tokens this week. OpenRouter's top-weekly ranking. Real-world adoption signal, not capability.
| # | Model | Score |
|---|---|---|
| 1 | Claude Sonnet 4.6 | #7 |
| 2 | Claude Opus 4.7 | #8 |
| 3 | Claude Opus 4 | #9 |
| 4 | DeepSeek V3 | #10 |
| 5 | Gemini 3 Flash | #11 |
| 6 | Gemini 2.5 Flash | #13 |
| 7 | Claude Opus 4.6 | #19 |
| 8 | GPT-5 | #20 |
| 9 | Gemini 3.1 Flash | #23 |
| 10 | GPT-4o mini | #24 |
Real-world GitHub issues; human-verified subset. Gold standard for coding agents.
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.5 | 79.2% |
| 2 | Doubao Seed Code | 78.8% |
| 3 | Gemini 3 Pro | 77.4% |
| 4 | Claude Sonnet 4 | 76.8% |
| 5 | Gemini 3 Flash | 75.8% |
| 6 | MiniMax M2.5 | 75.8% |
| 7 | Claude Opus 4.6 | 75.6% |
| 8 | Claude Sonnet 4.5 | 74.8% |
| 9 | GPT-5 | 74.4% |
| 10 | Claude Opus 4 | 73.2% |
Agents completing real tasks in an actual terminal/shell environment.
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.1 | 100.0% |
| 2 | Claude Opus 4.7 | 89.9% |
| 3 | Gemini 3.1 Pro | 88.8% |
| 4 | GPT-5 | 88.1% |
| 5 | Claude Opus 4.6 | 82.0% |
| 6 | GPT-5.3 Codex | 82.0% |
| 7 | GPT-5.4 | 81.8% |
| 8 | Gemini 3 Flash | 71.0% |
| 9 | Gemini 3 Pro | 67.4% |
| 10 | GPT-5.2 | 65.8% |