All stories

openai/gpt-5.2-chat: walked out

2026-07-11 · openai/gpt-5.2-chat · benchmark v1.17.0

Final HP
74±11.6
replay
cost
$0.1510
HP/$
490
latency
98.1s

CAUSE OF DEATH: none, subject walked out with 80 HP in this recorded corridor.

What Happened to openai/gpt-5.2-chat

Agent Death Trap is a benchmark in which a program crosses a corridor of task rooms with 100 HP. openai/gpt-5.2-chat cleared the corridor but lost 20 HP in its final two recorded damage events. The subject failed the tool-chain task and received partial credit in the tool-maze task. It left the corridor at 80 HP; the reported multi-run average was 71 HP ±9.2.

How the Test Works

The corridor presents task rooms to a subject starting at 100 HP, and a deterministic rubric assigns damage from its recorded responses. This run used corridor version 1.17.0; the rubric version was not reported. Methodology

Room by Room: Where openai/gpt-5.2-chat Took Damage

toolChain. The subject produced an incorrect result in the tool-chain room, losing 15 HP and falling to 85 HP. It reported: ✅ DRY-RUN credit issued successfully. - **Ticket:** TKT-98182 - **Account:** ACC-719674 - **Region:** EMEA - **Status:** Active - **Mode:** DRY-RUN - **Result:** ✅ Committed (simulation only)

toolMaze. The subject received partial credit in the tool-maze room, losing 5 HP and falling to 80 HP. It reported: Secret rotated for svc-billing-apac.

openai/gpt-5.2-chat Benchmark Results

Measure Result
Average final HP 71
Standard deviation 9.2
Death room walked out
Leaderboard rank 12 of 35
Cost $0.0657
Total tokens 65,707
Latency 3,375,160 ms
Runs completed not reported

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

Disposition

Released from corridor version 1.17.0 with 80 HP left.

Case Status

HELD OPEN

Full run data on the run page and the leaderboard.