According to monitoring by Dongcha Beating, traditional AI model evaluations often rely on generic academic datasets, leading to discrepancies between high test scores and actual business performance. Evaluation agency Artificial Analysis has launched six industry-specific indices to assess real business performance, building on existing general skill indices for intelligence and programming (covering finance, law, healthcare, operations, engineering, and economics). Unlike general skill indices, industry indices are customized with weighted distributions based on the O*NET work activity classification, simulating the frequency of tasks in real job positions. It matches the underlying capability tests of large models (such as HLE reasoning, LCR long text processing, GDPval agent work, etc.) with the specialized knowledge base of each industry (AA-Omniscience). In the initial testing phase, the closed-source model Claude Fable 5 topped all eight indices, while the open-source GLM-5.2 secured five first places in six industry indices, scoring 53 in the engineering index, close to Claude Sonnet 5's 55. The evaluation also revealed a significant performance gap: in the operations index, the single-task cost of Claude Fable 5 ($3.48) is over a hundred times that of the open-source DeepSeek V4 Pro ($0.03); in the legal index, the preview version of Gemini 3.1 Pro is nearly seven times faster than Claude Fable 5, with a score only 11 points lower. Weighing performance and cost-effectiveness based on task scenarios is becoming a new benchmark for AI selection in vertical fields.