Surfactant-Toolkit and SurfBench-200: an open computation toolkit and a benchmark of tool-augmented language-model reasoning in surfactant science

Language models are increasingly used as research assistants in chemistry, but their quantitative work insurfactantscience hasnotbeen measured.We introduce Surfactant-Toolkit, an MIT-licensed Python library and Model Context Protocol (MCP) server with 123 validated tools for micellization, mixed micelles, adsorption, emulsions, micelle structure and transport,andSurfBench-200,200graduate-levelquestionswhosereferenceanswersare computed by the toolkit and graded field by field by deterministic scripts. Without tools, five frontier models passed 52.5% to 66.0% of the questions; with the toolkit, the four models run in both conditions passed 92.0% to 98.5% (exact McNemar p < 0.0001 for each), and the open-weightDeepSeek (92.0%) outscored every modelwithouttools.Remaining failures shifted from wrong magnitudes tochoices ofmodelorconvention:toolslargely remove calculation errors, not errors of judgement. Asked the same questions again, every model changed some answers and at least once stated something the data did not support. Used for new findings, the toolkit and two augmented models produced five results on seven research problems that we could not find in the literature. Used for error finding, it showed that five recent papers each contain a value or conclusion that does not follow from their own printed data, from a reversed interaction-parameter sign to a physically impossible molecular area; two frontier models given the same values recomputed nearly all of these without tools but called an undetermined result contradicted. All questions, transcripts, graders and code are released, with SurfactantKit-Studio, a local application for researchers. 1. Introduction Surfactantsunderpindetergency,enhancedoilrecovery,emulsionpolymerisation, pharmaceutical solubilisation and personal-care formulation.[1] The routine computations of thefield,suchasextractingaCMC from tensiometryorconductivitydata,[2,3] mixedmicellethermodynamics,counterion binding,Gibbs and Szyszkowskiadsorption analysis