DeepSeek V4.1 Flash scores 73.69% at $0.07 per test; GPT-6 Sol scores 77.98% at $1.34 per test. The listed costs make GPT-6 Sol roughly 19 times more expensive per test, for a 4.29-point accuracy advantage.
First implication: Teams should validate both bug-finding accuracy and per-task cost on representative repositories before choosing a model.
Thesis
Benchmark utility for security agents depends on reading cost and task performance together, rather than treating a leaderboard rank as a deployment recommendation.
Observed facts
Mechanism
Paired accuracy and per-test cost figures show the price of the incremental benchmark performance, enabling teams to assess whether it matters for their workload.
Why now
CyberBench v1.1 presents these measures side by side for a current set of models, making the tradeoff legible to teams comparing security-task candidates.
Implications
Teams should validate both bug-finding accuracy and per-task cost on representative repositories before choosing a model.
Evidence boundary
The supplied source describes benchmark scores and costs only; it does not establish production performance, patch correctness in a team’s environment, or generalization beyond CyberBench.
Watch conditions
Whether subsequent evaluations report reproducible results across repository types and separate finding quality from patch quality.