AgenticRL-Bench Train Pool for TB21: leaderboard
AgenticRL-Bench Train Pool for TB21 — 19verifiable tasks×3attempts, headless one-shot8h, pass@1 solved/57. Three open-ended4×H200 tasks10–16h, means of3counted attempts in their native units. Elo is native BT within-task comparisons, not accuracy. Exact model/harness/effort retained; no best-harness pooling. Metric: Terminal-Bench2.1 pass@1 (%). Source: https://agenticrl-bench.env-lab.ai/leaderboard
Metric: Terminal-Bench2.1 pass@1 (%). Source: agenticrl-bench.env-lab.ai. Saturation forecast: Rough model projection: around 2026. 7 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | (Max) (AgenticRL run: harness=Claude Code) | 25.5 | #1 |
| 2 | Gemini 4 Carbon | (High) (AgenticRL run: harness=Antigravity) | 24.8 | #2 |
| 3 | Gemini 4 Argon | (High) (AgenticRL run: harness=Antigravity) | 24.1 | #12 |
| 4 | Claude Opus 5.5 | (Max) (AgenticRL run: harness=Claude Code) | 24 | #5 |
| 5 | GPT-6 Astra | (Max) (AgenticRL run: harness=Codex) | 22.4 | #6 |
| 6 | Grok 4.7 | (Xhigh) (AgenticRL run: harness=Grok Build) | 22.2 | #47 |
| 7 | Muse Wonder Full 1.0 | (Xhigh) (AgenticRL run: harness=Muse Code) | 19.1 |
Interactive version: theaggregate.ai/benchmark?slug=agenticrl-bench-train-pool-for-tb21 · How It Works · Data refreshed daily, snapshot 2026-11-16.