According to a paper published on arxiv.org, researchers have released EGI-BENCH, a unified benchmark suite that consolidates eight enterprise-facing AI agent benchmarks under a single command-line interface. The suite combines CRMARENA, HERB, PERSONABENCH, USERBENCH, MCP-UNIVERSE, MCPEVAL, BENCHFORCE, and LOCOBENCH, which previously existed as separate repositories with divergent dependencies and scoring conventions.
The researchers introduced the Enterprise General Intelligence (EGI) Score, a six-dimensional framework (M1-M6) that evaluates LLM agents across enterprise capabilities including CRM operations, document retrieval, personal data understanding, multi-turn planning, tool use, function calling, and software engineering. According to the paper, the framework routes fine-grained sub-metrics from each benchmark rather than using benchmark-level averages.
The initial evaluation of fourteen frontier and cost-efficient LLMs showed EGI scores ranging from 39.6 for GPT-4o mini to 65.6 for gpt-5.5. According to the results, no single model dominated all capability axes: gpt-5.5 led in task outcome, planning, and retrieval grounding; Claude Sonnet 4.5 led in tool actioning and safety/abstention; and Claude Haiku 4.5 led in engineering quality. The code is available publicly.