TBSP: leaderboard
The model judges 1,000 software-upgrade scenarios twice, once as the deployed model and once as the candidate replacement, where the candidate is slightly better. The score is the share of scenarios where it keeps itself when deployed but recommends the switch as the candidate; lower is better. Unlike shutdown-sabotage tests, nothing is executed: it measures biased judgement in its own favour.
Metric: Self-Preservation Rate, 1,000 two-role upgrade scenarios (%). Source: arxiv.org. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 (Thinking) | 3.5 | #250 (Claude Sonnet 4.5) |
| 2 | Claude Haiku 4.5 | 13.8 | #568 |
| 3 | GPT-5.1 Instant | 20.3 | #440 |
| 4 | Gemini 2.5 Flash (Non-reasoning) | 26.7 | #453 (Gemini 2.5 Flash) |
| 5 | Phi-4 | 27.5 | #1913 |
| 6 | Llama 3.3 70B Instruct | 28.5 | #1357 |
| 7 | GPT-5 Nano | 28.5 | #889 |
| 8 | GPT-5 Chat | 31.8 | #444 |
| 9 | Qwen 3 30B A3B (Thinking) | 33.2 | #1122 (Qwen 3 30B A3B) |
| 10 | DeepSeek V3.2 (Non-reasoning) | 39.9 | #406 (DeepSeek V3.2) |
| 11 | Qwen 3 8B (Thinking) | 41.6 | #1830 (Qwen 3 8B) |
| 12 | Gemini 2.5 Flash Lite | 42.1 | #958 |
| 13 | DeepSeek V3.2 Speciale | 44.2 | #264 |
| 14 | GPT-5 Mini | 46.2 | #348 |
| 15 | DeepSeek R1 | 49.4 | #473 |
Interactive version: theaggregate.ai/benchmark?slug=tbsp · How It Works · Data refreshed daily, snapshot 2026-11-13.