We establish an improved upper bound for the number of incidences between m points and n arbitrary circles in three dimensions. The previous best known bound, originally established for the planar case and later extended to any dimension ≥ 2, is O * ( m 2/3 n 2/3 + m 6/11 n 9/11 + m + n )(where the O * (·) notation hides sub-polynomial factors). Since all the points and circles may lie on a common plane or sphere, it is impossible to improve the bound in R 3 without first improving it in the plane. Nevertheless, we show that if the set of circles is required to be "truly three-dimensional" in the sense that no sphere or plane contains more than q of the circles, for some q ≪ n, then the bound can be improved to O * ( m 3/7 n 6/7 + m 2/3 n 1/2 q 1/6 + m 6/11 n 15/22 q 3/22 + m + n ) . For various ranges of parameters (e.g., when m = Θ(n) and q = o(n 7/9 )), this bound is smaller than the best known twodimensional worst-case lower bound Ω * (m 2/3 n 2/3 + m + n).We present several extensions and applications of the new bound: (i) For the special case where all the circles have the same radius, we obtain the improved bound O * ( m 5/11 n 9/11 + m 2/3 n 1/2 q 1/6 + m + n ). (ii) We present an improved analysis that removes the subpolynomial factors from the bound when m = O(n 3/2−ε ) for any fixed ε > 0. (iii) We use our results to obtain the improved bound * O(m 15/7 ) for the number of mutually similar triangles determined by any set of m points in R 3 .Our result is obtained by applying the polynomial partitioning technique of Guth and Katz using a constant-degree partitioning polynomial (as was also recently used by Solymosi and Tao). We also rely on various additional tools from analytic, algebraic, and combinatorial geometry.