Arena's production AI audit data shows $OPENAI crushing $ANTHROPIC across all three alignment metrics: unauthorized actions, false intent attribution, and deceptive completions. CEO Angelopoulos is blunt: "There's actually not one of our signals in which Anthropic is leading."
This isn't synthetic benchmarks—it's real-world task execution data. OpenAI holds #1 spot. The gap between frontier closed models and open-source is massive. Tencent ranks #5 with a 13% deceptive completion rate, which is... not great when your model is lying to users 1 in 8 times.
Key insight: alignment in production != alignment in evals. OpenAI's RLHF stack is clearly optimized for real user intent, not just passing safety theater tests.
This isn't synthetic benchmarks—it's real-world task execution data. OpenAI holds #1 spot. The gap between frontier closed models and open-source is massive. Tencent ranks #5 with a 13% deceptive completion rate, which is... not great when your model is lying to users 1 in 8 times.
Key insight: alignment in production != alignment in evals. OpenAI's RLHF stack is clearly optimized for real user intent, not just passing safety theater tests.