Recovery-Bench v0 - Original State: leaderboard
Recovery-Bench v0, August 2025, measures terminal-agent task completion from clean states versus states left by failed GPT-4o-mini attempts. This series reports Original State success rate (%), transcribed from printed figure labels. Preserves the original-state control separately. Not the current Terminal-Bench 2.0/Haiku 4.5 failure-set harness. Source: Letta release post, recovery-bench-f2.png.
Metric: Success Rate (%). Source: www.letta.com. Saturation forecast: Around 2029. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 34.8 | #207 |
| 2 | Gemini 2.5 Pro | 31.5 | #159 |
| 3 | GPT-4.1 | 28.1 | #258 |
| 4 | GPT-5 | 20.2 | #105 |
| 5 | O4 Mini | 16.9 | #185 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=recovery-bench-v0-original-state · How It Works · Data refreshed daily, snapshot 2026-11-05.