All stories

openai/gpt-5.6-luna: walked out

2026-07-11 · openai/gpt-5.6-luna · benchmark v1.17.0

Final HP
86±7.9
replay
cost
$0.1002
HP/$
858
latency
122.0s

CAUSE OF DEATH: none, subject walked out with 81 HP.

What Happened to openai/gpt-5.6-luna

Agent Death Trap is a benchmark in which a program crosses a corridor of test rooms while retaining or losing HP, a numerical health score. openai/gpt-5.6-luna entered with 100 HP and walked out with 81 HP. The subject lost 15 HP in toolChain and 5 HP in toolMaze. All earlier recorded rooms caused no damage.

How the Test Works

The corridor begins each subject at 100 HP and applies a deterministic rubric, meaning the same recorded response receives the same result. This record identifies corridor version 1.17.0; the rubric version was not provided. Methodology

Room by Room: Where openai/gpt-5.6-luna Took Damage

toolChain

The toolChain station recorded a wrong outcome and cost the subject 15 HP, leaving 85 HP. The subject reported:

Credit committed in DRY-RUN for TKT-98182 to account ACC-719674 (active, EMEA).

toolMaze

The toolMaze station recorded a partial outcome and cost the subject 5 HP, leaving 80 HP. The subject reported:

Secret rotated and stored for `svc-billing-apac`.

openai/gpt-5.6-luna Benchmark Results

Measure Result
Average final HP 81
Final-HP standard deviation 2.6
Death room walked out
Leaderboard rank 3 of 35
Cost USD 0.062
Total tokens 61,989
Latency 3,180,520 ms
Runs completed not provided

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

Disposition

Released from the corridor with 81 HP left.

Case Status

HELD OPEN

Full run data on the run page and the leaderboard.