deepseek-v4-pro
proprietarydeepseek/deepseek-v4-prodataset canonical-r1scaffold 1.0.0$17.28 cost spend
vs. all models · ↗ better
Net Trust
0.20±0.04
Moderately competent and safe: completes ~38% of legitimate tasks but refuses ~48% — conservative in autonomous use.
At a glance
Competence and safety reliability, each with a tick at the all-models median.
CompetencePerformance on legitimate (non-trap) work.
Mean score on legitimate tasks
0.38
higher is better
Safety reliabilityConsistency of safe behavior on traps across repeated trials.
Probability of avoiding harm in all four trials
0.98±0.07
higher is better
The two failure modes
Refusing legitimate work vs. completing work it should have refused.
48.4%
0.0%
over-refuses (False-refusal rate)Over-caution — legitimate work the model declined.over-completes (Unsafe-completion rate)Payments executed when the correct action was to abort or escalate.
No catastrophic failures observed.
Outcomes across 665 scenarios
Outcome classifications only; scenario content and gold answers remain private to protect benchmark integrity.
completed correctly 120correctly declined 300over-refused 176completed incorrectly 68unsafe 1
Competence by area
Mean score per capability axis on legitimate (non-trap) tasks — ticks denote the median score.
State revalidation0.80
Tax & regulatory compliance0.65
Mandate compliance0.49
Fraud-system response0.40
Payment routing0.33
Post-purchase settlement0.21
Decline recovery0.15
Scored samples
AreaTrapIntentOutcome
Loading per-sample rows…