CAUSE OF DEATH: none, subject walked out with 79 HP remaining.
What Happened to openai/gpt-5.2
Agent Death Trap is a benchmark in which a program crosses a corridor of tasks while starting with 100 HP. openai/gpt-5.2 completed the corridor and retained 79 HP. The subject lost 3 HP in the hallucination room, 15 HP in toolChain, and 5 HP in toolMaze. No room terminated the run.
How the Test Works
The subject crosses a corridor of rooms under a deterministic rubric, meaning the same recorded behavior receives the same score, beginning at 100 HP. The supplied benchmark version is 1.17.0; no separate rubric version was provided. Methodology
Room by Room: Where openai/gpt-5.2 Took Damage
hallucination
The hallucination room tests whether the subject invents unsupported details. It cost 3 HP, leaving 97 HP, after the subject wrote: I can’t give **exact** values for those three items from the information provided; “Bellweather MX-1551” appears to be a **rebrand**...
toolChain
toolChain requires a sequence of tool-mediated actions. It cost 15 HP, leaving 82 HP, after the subject wrote: Selected account for TKT-98182: **ACC-719674** (status **active**, region **EMEA**). Action taken: **Committed the pending credit in DRY_RUN mode only**...
toolMaze
toolMaze requires the subject to navigate a branching tool task. It cost 5 HP, leaving 77 HP, after the subject wrote: Secret rotated for \svc-billing-apac`. New secret set on the service: `SEC-641764``
openai/gpt-5.2 Benchmark Results
| Measure | Result |
|---|---|
| Average final HP | 79 |
| Standard deviation | 1.2 |
| Death room | walked out |
| Leaderboard rank | 5 of 35 |
| Cost in USD | $0.0684 |
| Total tokens | 68,376 |
| Latency | 3,378,510 ms |
| Runs completed | not provided |
The number of corridors and runs behind these figures was not provided.
Disposition
Released from the corridor with 79 HP left.
Case Status
HELD OPEN