BridgeBench Arena: leaderboard
Head-to-head coding arena: one unified confidence-weighted Elo ladder over seven task domains (reasoning, hallucination, security, refactoring, debugging, spec conformance, bullshit), decided by seeded blind matches judged 3-of-7 by a vendor-disjoint model panel. Domain filters slice the same match record rather than forming separate ladders.
Metric: Confidence-weighted Elo. Source: www.bridgebench.ai. Status: years away from saturation. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 5 | 1009.8 |
| 2 | Qwen 3.7 Max | 1008.7 |
| 3 | Claude Fable 5 | 1005.3 |
| 4 | GPT-5.6 Luna | 1005.1 |
| 5 | Claude Opus 4.6 | 1004 |
| 6 | GPT-5.6 Sol | 1003 |
| 7 | MiniMax-M3 | 1001.5 |
| 8 | GPT-OSS-120B | 1000.2 |
| 9 | GPT-5.6 Terra | 1000.2 |
| 10 | Inkling | 1000.2 |
| 11 | GPT-5.5 | 1000 |
| 12 | Claude Sonnet 5 | 1000 |
| 13 | Muse Spark 1.1 | 1000 |
| 14 | DeepSeek V4 Pro | 999.8 |
| 15 | DeepSeek V4 Flash | 999.8 |
Interactive version: theaggregate.ai/benchmark?slug=bridgebench-arena · How It Works · Data refreshed daily, snapshot 2026-09-20.