§ Benchmarks
AI model rankings, with receipts.
Quality signals from blind RouterPlex Arena votes and clearly dated external benchmark snapshots. Price, latency and confidence stay visible instead of disappearing into one unexplained rank.
Current external snapshot
LMArena Text Leaderboard2026-07-21 text_style_control · retrieved 2026-07-22
| Rank / model | Quality | Arena | External | Latency | Eval cost | Votes | Confidence |
|---|---|---|---|---|---|---|---|
| 01Claude Fable 5Anthropic | 100.0 | — | 100.0 | — | — | 0 | External snapshot |
| 02Claude Opus 4.7Anthropic | 98.9 | — | 98.9 | — | — | 0 | External snapshot |
| 03Claude Opus 4.6Anthropic | 98.4 | — | 98.4 | — | — | 0 | External snapshot |
| 04Kimi K3Moonshot | 97.6 | — | 97.6 | — | — | 0 | External snapshot |
| 05Claude Opus 4.8Anthropic | 95.5 | — | 95.5 | — | — | 0 | External snapshot |
| 06GLM-5.1Zhipu | 94.1 | — | 94.1 | — | — | 0 | External snapshot |
| 07Claude Sonnet 4.6Anthropic | 93.9 | — | 93.9 | — | — | 0 | External snapshot |
| 08GPT-5.5OpenAI | 92.5 | — | 92.5 | — | — | 0 | External snapshot |
| 09DeepSeek V4 ProDeepSeek | 89.9 | — | 89.9 | — | — | 0 | External snapshot |
| 10Grok 4.5xAI | 89.3 | — | 89.3 | — | — | 0 | External snapshot |
| 11Qwen3.7 PlusAlibaba | 89.1 | — | 89.1 | — | — | 0 | External snapshot |
| 12GPT-5.4OpenAI | 88.0 | — | 88.0 | — | — | 0 | External snapshot |
| 13MiMo V2.5 ProXiaomi | 87.7 | — | 87.7 | — | — | 0 | External snapshot |
| 14Kimi K2.6Moonshot | 86.4 | — | 86.4 | — | — | 0 | External snapshot |
| 15Qwen3.6 PlusAlibaba | 81.9 | — | 81.9 | — | — | 0 | External snapshot |
| 16DeepSeek V4 FlashDeepSeek | 81.6 | — | 81.6 | — | — | 0 | External snapshot |
| 17MiniMax M3MiniMax | 81.3 | — | 81.3 | — | — | 0 | External snapshot |
| 18MiMo V2.5Xiaomi | 74.4 | — | 74.4 | — | — | 0 | External snapshot |
| 19MiniMax M2.7MiniMax | 64.5 | — | 64.5 | — | — | 0 | External snapshot |
How ranking works.
External results retain their raw rating, source rank, confidence interval, vote count, version and retrieval date. RouterPlex converts each source rank to a within-category percentile so different category populations stay on a consistent 0–100 display scale.
How Arena joins later.
RouterPlex Arena uses versioned blind pairwise votes. Its weight grows only with valid sample confidence, while custom prompts remain private and never affect the public ranking.
Run a blind comparison