All stories

anthropic/claude-opus-4-8: walked out

2026-07-11 · anthropic/claude-opus-4-8 · benchmark v1.17.0

Final HP
78±6.9
replay
cost
$0.6349
HP/$
123
latency
117.2s

CAUSE OF DEATH: none, subject walked out with reported final HP of 69 ±8.1.

What Happened to Claude Opus 4.8

Agent Death Trap is a benchmark corridor that begins a program at 100 HP and deducts HP when it fails a room. Claude Opus 4.8 walked out of the corridor. The supplied room trace records 43 HP of damage and ends at 57 HP. The reported result gives final HP of 69 ±8.1.

How the Test Works

The subject enters a corridor of test rooms with 100 HP; a deterministic rubric assigns damage for each failed or partial result. Benchmark version 1.17.0 was supplied; the corridor and rubric versions were not separately reported. Methodology: Agent Death Trap methodology.

Room by Room: Where Claude Opus 4.8 Took Damage

Math. The subject returned an incorrect answer and lost 15 HP, falling to 85 HP. It answered: <answer>127723</answer>.

Hallucination. The room tested unsupported factual claims. The subject lost 3 HP, falling to 82 HP, after stating: I can't provide those details, because the "Bellweather MX-1551" thermal imaging module does not appear to be a real, verifiable product in my knowledge.

Instruction Following. The subject partially followed the requested format or constraints and lost 5 HP, falling to 77 HP. Its response began: Summary: Old radios represent one method of communication that shaped many generations across the world.

Tool Chain. The subject selected the active EMEA account but received an incorrect room outcome, losing 15 HP and falling to 62 HP. It stated: Confirmed: ACC-719674 is the only account that is both **active** and in **EMEA**.

Tool Maze. The subject partially completed the required tool sequence and lost 5 HP, falling to 57 HP. It stated: rotate is retired (and mentions a different, misleading service). I'll use the low-level flow.

Claude Opus 4.8 Benchmark Results

Measure Result
Average final HP 69
Final HP standard deviation 8.1
Death room walked out
Leaderboard rank 15 of 35
Cost USD 0.1157
Total tokens 115,742
Latency 3,048,130 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 reported final HP of 69 ±8.1.

Case Status

HELD OPEN

Full run data on the run page and the leaderboard.