Breaking01 / AI
EGI-BENCH puts eight enterprise agent benchmarks in one tool
The authors report that no model leads on every capability axis, but the abstract gives no deep-mode scores.

Zhiwei Liu and colleagues reported in EGI-BENCH, a preprint submitted to arXiv on 19 Oct 2026 and not peer-reviewed, that no single language model leads on every capability axis of a combined enterprise agent benchmark suite (n=14, quick mode); the paper page presents no independent replication.
The authors said they combined eight enterprise agent benchmarks, from CRMARENA to LOCOBENCH, behind one command-line tool. Coverage spans customer relationship management (CRM) tasks, retrieval across heterogeneous documents, tool calling, long-horizon research and software development with long context. Runs are split into a quick mode and a deep mode with different time budgets, and the reported scores come from quick mode.
The Enterprise General Intelligence (EGI) score, which the authors define, has six dimensions, labelled M1 to M6. Composite EGI scores range from 39.6 for GPT-4o mini to 65.6 for gpt-5.5 (n=14, quick mode), the authors reported.
By capability, according to the authors, gpt-5.5 leads on task outcome, planning and retrieval grounding (n=14, quick mode). Claude Sonnet 4.5 leads on tool execution and on safety and abstention, and Claude Haiku 4.5 leads on engineering quality.
The scores in the abstract are limited to quick-mode results for models in the GPT, Claude and Gemini families. The authors said they have released their code, and the paper page shows no replication by outside researchers.
Existing evaluations are hard to compare across benchmarks, the authors argued, because each repository has its own runners, dependencies and scoring rules. The abstract says the EGI score does not use benchmark-level averages and instead assigns detailed metrics from each benchmark to each dimension.
The abstract does not report deep-mode scores or scores for the individual M1-M6 dimensions.