Benchmarks · Science & research · Coding · Agents & computer use
FrontierRSI
Can an agent do the research and engineering that improves AI? 27 open-ended tasks on real GPUs, from training methods to serving and GPU kernels, two trials each. A trial scores 1 for matching the best result any model reached on that task and 0 at 10% worse.
9 models · 9 results · data as of 2026-11-18 · frontier-rsi.org ↗
Leaderboard
Opus 5.5 max
Heavy line: best score at each price. Thin lines join one model's effort levels. Cost uses each model's most subsidized plan, or API price with no plan data.
More from this benchmark
5 published measuresScore vs any published measure5 measures: sub-scores, splits, counts
Breakdown
darker is better within each column · click a column to sort| Model | ||||
|---|---|---|---|---|
| Opus 5.5 max | 44.1 | 19.0 | 27.7 | 46.1 |
| Fable 5.2 max | 41.4 | 8.21 | 66.3 | 31.7 |
| Gemini 4 Carbon high | 26.9 | 4.32 | 30.7 | 53.6 |
| Gemini 4 Argon high | 26.8 | 13.5 | 22.9 | 31.3 |
| GPT-6.1 Sol max | 44.1 | 0 | 40.1 | 23.7 |
| GPT-6 Astra max | 25.4 | 12.5 | 18.7 | 27.1 |
| Kimi K3.1 max | 15.7 | 13.4 | 8.25 | 15.6 |
| Muse Wonder max | 14.6 | 2.05 | 20.0 | 14.9 |
| Grok 4.7 xhigh | 9.25 | 11.8 | 4.37 | 8.66 |
All results
9 rows| # | Model | Effort | FrontierRSI | Subscription $ / task | API $ / task | Harness |
|---|---|---|---|---|---|---|
| 1 | Opus 5.5 | max | 34.3% | $0.27 | $21 | Claude Code |
| 2 | Fable 5.2 | max | 33.0% | $1.37 | $37 | Claude Code |
| 3 | Gemini 4 Carbon | high | 29.8% | Antigravity CLI | ||
| 4 | Gemini 4 Argon | high | 23.6% | $4.91 | $7.37 | Antigravity CLI |
| 5 | GPT-6.1 Sol | max | 23.5% | $0.33 | $9.84 | Codex |
| 6 | GPT-6 Astra | max | 20.9% | $1.87 | $52 | Codex |
| 7 | Kimi K3.1 | max | 13.6% | Claude Code | ||
| 8 | Muse Wonder | max | 12.0% | $3.24 | $4.86 | Muse Code |
| 9 | Grok 4.7 | xhigh | 8.9% | Grok Build |