Benchmarks · Science & research · Coding
FrontierRSI: Pre-training
FrontierRSI's pre-training tasks: improve optimizers, objectives, architectures and draft-model training, scored on held-out loss or latency against the best result any model reached.
9 models · 9 results · data as of 2026-11-18 · frontier-rsi.org ↗
Leaderboard
GPT-6.1 Sol max
All results
9 rows| # | Model | Effort | Score | Harness |
|---|---|---|---|---|
| 1 | GPT-6.1 Sol | max | 44.1% | Codex |
| 2 | Opus 5.5 | max | 44.1% | Claude Code |
| 3 | Fable 5.2 | max | 41.4% | Claude Code |
| 4 | Gemini 4 Carbon | high | 26.9% | Antigravity CLI |
| 5 | Gemini 4 Argon | high | 26.8% | Antigravity CLI |
| 6 | GPT-6 Astra | max | 25.4% | Codex |
| 7 | Kimi K3.1 | max | 15.7% | Claude Code |
| 8 | Muse Wonder | max | 14.6% | Muse Code |
| 9 | Grok 4.7 | xhigh | 9.3% | Grok Build |