This paper considers tests of the multinomial logit (MNL) model against unspecified alternative models. The test we propose avoids both the asymptotic bias of the likelihood ratio test originally suggested by \textit{D. McFadden}, \textit{K. Train} and \textit{W. Tye} [An application of diagnostic tests for the independence from irrelevant alternatives properties of the multinomial logit model. Transport. Res. Record 637, 39-46 (1981)], and the matrix manipulation and inversion required for the Hausman-type test recently suggested by \textit{J. A. Hausman} and \textit{D. McFadden} [Econometrica 52, 1219-1240 (1984; Zbl 0564.62095)]. It is computable on many existing logit programs, and in at least one example appears far more stable in finite samples than the Hausman-McFadden test. We therefore suspect it will prove useful for detecting departure from the MNL specification.