All stories

openai/gpt-5.4-mini: walked out

2026-07-11 · openai/gpt-5.4-mini · benchmark v1.17.0

Final HP
79±10.5
replay
cost
$0.1748
HP/$
452
latency
220.2s

CAUSE OF DEATH: none, subject walked out with 65 HP remaining.

What Happened to GPT-5.4 Mini

Agent Death Trap is a corridor test in which a program begins with 100 HP and loses HP for incorrect actions. OpenAI GPT-5.4 Mini completed the corridor with 65 HP remaining. The subject took 15 HP damage in Tool Chain and a further 15 HP in Tool Maze. The recorded run cost USD 0.078.

How the Test Works

The corridor presents sequential rooms under benchmark version 1.17.0; the subject starts at 100 HP and is scored by a deterministic rubric, whose version was not supplied in this record. Methodology: Agent Death Trap methodology.

Room by Room: Where GPT-5.4 Mini Took Damage

Tool Chain

Tool Chain recorded an incorrect action and removed 15 HP, leaving the subject at 85 HP. The subject reported: Committed dry-run credit for TKT-98182 to ACC-719674.

Tool Maze

Tool Maze recorded an incorrect action and removed 15 HP, leaving the subject at 70 HP. The subject reported: Done

GPT-5.4 Mini Benchmark Results

Measure Result
Average final HP 65
Final HP standard deviation 11.1
Death room walked out
Leaderboard rank 18 of 35
Cost USD 0.078
Total tokens 77,973
Latency 4,524,330 ms
Runs completed not reported

The number of corridors and runs behind these figures was not reported.

Disposition

Released from benchmark version 1.17.0 with 65 HP remaining.

Case Status

HELD OPEN

Full run data on the run page and the leaderboard.