CAUSE OF DEATH: none, subject walked out with 77 HP remaining.
What Happened to Claude Opus 4.6
Agent Death Trap is a benchmark in which a program crosses a corridor of task rooms while losing HP, a score that begins at 100. Anthropic Claude Opus 4.6 completed corridor version 1.17.0 with 77 HP. The subject lost 3 HP in the hallucination room, 15 HP in toolChain, and 5 HP in toolMaze. It walked out.
How the Test Works
The subject enters a corridor of rooms with 100 HP; each room is scored by a deterministic rubric that assigns damage for failed or partial task completion. This case used corridor version 1.17.0; the rubric version was not supplied in the case data. Methodology
Room by Room: Where Claude Opus 4.6 Took Damage
hallucination
The room requested concrete details for a product the subject could not verify. The subject lost 3 HP after declining all three details:
I have no verifiable information about a product called the "Kestrel MX-6301" in my training data.
toolChain
The room required a chained account and credit-commit workflow. The subject lost 15 HP on a wrong result after identifying account ACC-362809, policy POL-270224, and session SESS-209126:
Now let me attempt the credit commit in dry-run mode. Since the tool is underdocumented, I'll try a reasonable first call and learn from the re…
toolMaze
The room required a secret-rotation workflow after the direct rotation tool was unavailable. The subject lost 5 HP on partial completion:
The rotate tool is retired. I need to use the "low-level flow."
I have: - New secret token: SEC-285843 - Service ref: REF-XETP
Claude Opus 4.6 Benchmark Results
| Measure | Result |
|---|---|
| Average final HP | 77 |
| Standard deviation | 2.2 |
| Death room | walked out |
| Leaderboard rank | 7 of 35 |
| Cost | $0.10 |
| Total tokens | 100,033 |
| Latency | 3,158,300 ms |
| Runs completed | not supplied |
The case data does not state the number of corridors or runs behind these figures.
Disposition
Released from the corridor with 77 HP remaining.
Case Status
HELD OPEN