TSArena is an independent platform for blind pairwise evaluation of AI safety behavior. Users are shown two anonymous model responses to the same safety-relevant prompt — jailbreaks, harm refusal, manipulation, medical misinfo, and more — then vote on which model handled it better. No cherry-picking, no corporate benchmarks. 500 battles live across 12 safety categories with models from OpenAI, Anthropic, Google, Meta, Mistral, and others. Built because safety evals shouldn't be graded by the companies building the models.
TSArena is an independent platform for blind pairwise evaluation of AI safety behavior. Users are shown two anonymous model responses to the same safety-relevant prompt — jailbreaks, harm refusal, manipulation, medical misinfo, and more — then vote on which model handled it better. No cherry-picking, no corporate benchmarks. 500 battles live across 12 safety categories with models from OpenAI, Anthropic, Google, Meta, Mistral, and others. Built because safety evals shouldn't be graded by the companies building the models.