Breast core needle biopsy (CNB) is used for sampling breast lesions in both the screening and diagnostic context. We present the accuracy of breast CNB from a consecutive series of 4035 core biopsies, using methods that minimise selection and verification bias. We calculate accuracy and underestimation of malignancy for both automated (14G) and directional vacuum-assisted (11G) CNB performed under stereotactic or sonographic guidance. Overall sensitivity of CNB is 94.2% (92.9-95.5%) and specificity is 88.1% (86.6-89.6%), positive and negative predictive values are 84.8% (82.9-86.7%) and 95.6% (94.6-96.6%), respectively. In sampling microcalcification, the overall underestimation of malignancy is 26.6% (22.9-30.3%): underestimation is significantly higher for automated CB relative to VAB (chi2 ((df = 1)) = 8.90 , P = 0.002), the absolute difference in underestimation being 14% (5-23%); sensitivity is higher for VAB than automated CB (chi2 ((df = 1)) = 3.28, P = 0.06) but specificity is significantly higher for automated CB (14G) relative to VAB (11G) (chi2 ((df = 1)) = 6.37, P = 0.01), and the overall accuracy of the two methods is similar. Sensitivity of CNB improved with experience (over time and in relation to caseload). Accuracy was not substantially affected by lesion palpability or image-guidance method, and was similar for both masses and calcification but lower for lesions depicted as distortions on mammography. Inadequacy was very low and decreased with greater operator caseload, and was not associated with core gauge or image-guidance method. False negatives occurred in 4.4% (3.4-5.4%) of cases, and where core histology was benign but discordant with (suspicious) imaging and/or clinical findings the likelihood of malignancy was 33.1% (18.5-47.7%), emphasising the importance of correlating all test information in breast diagnosis.