New EGI-BENCH Suite Scores Claude Models Against GPT and Gemini

Researchers from Salesforce AI Research published EGI-BENCH, a benchmark suite built to answer a narrower, more practical question than most agent leaderboards attempt: which model should a company actually run for a given enterprise workflow. The suite folds eight existing agent benchmarks, CRMARENA, HERB, PERSONABENCH, USERBENCH, MCP-UNIVERSE, MCPEVAL, BENCHFORCE and LOCOBENCH, behind one command line interface, replacing the separate runners, dependencies and scoring conventions each had previously used. On top of that consolidated suite, the authors define an Enterprise General Intelligence score built from six dimensions, labeled M1 through M6, each populated by fine grained sub-metrics routed from the underlying benchmarks rather than by averaging whole benchmark scores together.

The paper reports results for fourteen models spanning the GPT, Claude and Gemini families, run under the suite’s quick, runtime budgeted mode. The authors found that no single model wins across every capability axis. GPT-5.5 leads on task outcome, planning and retrieval grounding, while Claude Sonnet 4.5 leads on tool actioning and on safety and abstention, and Claude Haiku 4.5 leads on engineering quality. That spread is the paper’s central claim: an aggregate ranking collapses distinctions that matter once a buyer is choosing a model for one specific job rather than for everything at once.

The numbers

Top-line EGI scores across the fourteen tested models ranged from 39.6, posted by GPT-4o mini, up to 65.6, posted by gpt-5.5. The suite reports each model’s performance separately on the six underlying dimensions rather than folding them into one number, which is how the paper surfaces Claude Sonnet 4.5’s lead on tool actioning and safety abstention even though it does not top the overall ranking.

For teams building agentic workflows around CRM operations, document retrieval or long-context software engineering, the paper’s own framing is a caution against buying on a single leaderboard position. A model that trails on an aggregate score can still be the better choice for a workflow that depends heavily on the one dimension it leads, and the suite’s unified runner is built specifically so that comparison can be made without rerunning eight separate benchmark repositories by hand.