RSI-Exam: leaderboard
rsi-exam.ai: can an agent improve the machine-learning system it is handed? 88 recursive-self-improvement tasks, each normalised so the shipped baseline scores 0, a frontier-calibrated reference 0.6 and a proven upper bound 1; the published number is the mean over the hidden test set. Each model runs in its own vendor CLI at its top effort setting.
Metric: Mean hidden-set score. Source: rsi-exam.ai. Saturation forecast: Around March 2028. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 0.54 | #3 (Claude Opus 5.5) |
| 2 | GPT-6 Astra (Max) | 0.51 | #6 (GPT-6 Astra) |
| 3 | Claude Fable 5.1 (Max) | 0.48 | #9 (Claude Fable 5.1) |
| 4 | Claude Opus 5 (Max) | 0.46 | #12 (Claude Opus 5) |
| 5 | GPT-5.6 Sol (Max) | 0.43 | #19 (GPT-5.6 Sol) |
| 6 | GLM-5.3 (Max) | 0.4 | #39 (GLM-5.3) |
| 7 | Muse Spark 1.3 (Max) | 0.39 | #24 (Muse Spark 1.3) |
| 8 | Grok 4.7 (xHigh) | 0.38 | #25 (Grok 4.7) |
| 9 | Kimi K3 (Max) | 0.38 | #30 (Kimi K3) |
| 10 | Grok 4.6 (xHigh) | 0.37 | #26 (Grok 4.6) |
| 11 | Gemini 3.8 Flash (High) | 0.34 | #21 (Gemini 3.8 Flash) |
| 12 | GPT-5.5 (xHigh) | 0.33 | #31 (GPT-5.5) |
| 13 | DeepSeek V4 Pro (Max) | 0.32 | #102 (DeepSeek V4 Pro) |
| 14 | Qwen 3.8 Max (xHigh) | 0.32 | #35 (Qwen 3.8 Max) |
| 15 | Gemini 3.7 Flash (High) | 0.31 | #23 (Gemini 3.7 Flash) |
No result here: #7 GPT-6.1 Sol, #8 Gemini 4 Argon, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=rsi-exam · How It Works · Data refreshed daily, snapshot 2026-10-20.