The verdict
Loading the verdict…
CyBT-CTF asks 59 blinded investigation questions over live Splunk data.
Accepted runs
Matched Claude Code runs
Across versions
Compare within each family
White outlines mark 5.5; dotted outlines mark reported scores.
Claude Code · Opus
Values and conditions
Claude Code · Sonnet
Values and conditions
Claude Code · Haiku
Values and conditions
CyBT-CTF leaderboard
Claude 5.5 and its strongest predecessors
Previous Opus and Sonnet; latest and best available reference models.
Cost and speed
What each solved answer costs
Refusals and abstentions
Why a run came back without an answer
A refusal declines the task. An abstention reports that the model cannot determine the answer after investigating.
Benchmark validity