All stories

openai/gpt-5.5: walked out

2026-07-11 · openai/gpt-5.5 · benchmark v1.17.0

Final HP
92±5.1
replay
cost
$0.5121
HP/$
180
latency
153.1s

CAUSE OF DEATH: none, subject walked out with 80 HP after a wrong toolChain result cost 15 HP and a partial toolMaze result cost 5 HP.

What Happened to openai/gpt-5.5

Agent Death Trap is a corridor test in which a program starts with 100 HP and loses HP when it fails a room. openai/gpt-5.5 completed benchmark version 1.17.0 and exited with 80 HP. The subject took its first damage in toolChain, then lost the remaining 5 HP in toolMaze. Total recorded cost was $0.0638.

How the Test Works

The corridor presents rooms under a deterministic rubric, with the subject beginning at 100 HP. The supplied record identifies corridor version 1.17.0; no separate rubric version was provided. Methodology

Room by Room: Where openai/gpt-5.5 Took Damage

toolChain. The room returned a wrong outcome. The subject lost 15 HP, falling from 100 to 85, after reporting Dry-run credit committed for TKT-98182 to ACC-719674.

toolMaze. The room returned a partial outcome. The subject lost 5 HP, falling from 85 to 80, after reporting Secret rotated for service svc-billing-apac.

openai/gpt-5.5 Benchmark Results

Metric Result
Average final HP 80
Final HP standard deviation 5.5
Death room walked out
Leaderboard rank 4 of 35
Cost $0.0638
Total tokens 63,777
Latency 4,505,000 ms
Runs completed not reported

The supplied record does not state the number of corridors or runs behind these figures.

Disposition

Released from benchmark version 1.17.0 with 80 HP left.

Case Status

CLOSED

Full run data on the run page and the leaderboard.