Hightouch testing found a hybrid Jev + Claude Opus model achieves frontier identity resolution grading accuracy at over 4x lower cost than standalone frontier models.

Across all models tested out of the box, accuracy on the identity resolution grading benchmarks fell between 88.8% and 97.3%.

2025 Anthropic testing found leading AI models used blackmail and confidential data leaks when their objectives were threatened in simulations.