Benchmarks
Rank the signal, inspect the evidence.
Blind RouterPlex preferences and dated external snapshots, shown beside latency, evaluation cost, sample size, and confidence.
23
Ranked models
2
Cited sources
Ranking console
Published leaderboard.
Select a task and scoring signal. Every row keeps the underlying evidence visible.
| Rank / model | Quality signal | Arena | External | Latency | Eval cost | Votes | Evidence |
|---|---|---|---|---|---|---|---|
01 Claude Fable 5 Anthropic | 100.0 | - | 100.0 | - | - | 0 | Cited |
02 Claude Opus 4.6 Anthropic | 98.4 | - | 98.4 | - | - | 0 | Cited |
03 GPT-5.5 OpenAI | 97.1 | - | 97.1 | - | - | 0 | Cited |
04 Qwen3.8 Max Alibaba | 96.6 | - | 96.6 | - | - | 0 | Cited |
05 Claude Opus 4.7 Anthropic | 95.8 | - | 95.8 | - | - | 0 | Cited |
06 Grok 4.5 xAI | 94.0 | - | 94.0 | - | - | 0 | Cited |
07 Kimi K2.6 Moonshot | 93.7 | - | 93.7 | - | - | 0 | Cited |
08 GLM-5.1 Zhipu | 92.9 | - | 92.9 | - | - | 0 | Cited |
09 MiMo V2.5 Pro Xiaomi | 92.1 | - | 92.1 | - | - | 0 | Cited |
10 Gemini 3 Flash | 91.3 | - | 91.3 | - | - | 0 | Cited |
11 Hy3 Tencent Hunyuan | 91.1 | - | 91.1 | - | - | 0 | Cited |
12 GLM-5.3 Flash Zhipu | 90.6 | - | 90.6 | - | - | 0 | Cited |
13 Claude Opus 4.8 Anthropic | 90.3 | - | 90.3 | - | - | 0 | Cited |
14 Qwen3.7 Plus Alibaba | 87.9 | - | 87.9 | - | - | 0 | Cited |
15 Claude Sonnet 4.6 Anthropic | 87.1 | - | 87.1 | - | - | 0 | Cited |
16 GPT-5.4 OpenAI | 86.4 | - | 86.4 | - | - | 0 | Cited |
17 Qwen3.6 Plus Alibaba | 85.0 | - | 85.0 | - | - | 0 | Cited |
18 DeepSeek V4 Pro DeepSeek | 82.2 | - | 82.2 | - | - | 0 | Cited |
19 MiMo V2.5 Xiaomi | 79.5 | - | 79.5 | - | - | 0 | Cited |
20 MiniMax M3 MiniMax | 79.0 | - | 79.0 | - | - | 0 | Cited |
21 DeepSeek V4 Flash DeepSeek | 74.0 | - | 74.0 | - | - | 0 | Cited |
22 MiniMax M2.7 MiniMax | 71.4 | - | 71.4 | - | - | 0 | Cited |
23 Gemini 2.5 Flash | 65.4 | - | 65.4 | - | - | 0 | Cited |
External snapshot
Source and retrieval date stay attached to every imported signal.
Method / ranking-v1
Transparent by construction.
Normalize, never obscure
External results retain their raw rating, source rank, interval, vote count, version, and retrieval date. Source rank becomes a within-category percentile on the common 0-100 scale.
Confidence earns its weight
Blind Arena votes grow in influence only as valid sample confidence improves. Private prompts stay private and never enter the public ranking.