Recently, anchor-free detectors have shown great potential to outperform anchor-based detectors in terms of both accuracy and speed. In this work, we aim at finding a new balance of speed and accuracy for anchor-free detectors. Two questions are studied: 1) how to make the anchor-free detection head better? 2) how to utilize the power of feature pyramid better? We identify attention bias and feature selection as the main issues for these two questions respectively. We propose to address these issues with a novel training strategy that has two soften optimization techniques, i.e. soft-weighted anchor points and soft-selected pyramid levels. To evaluate the effectiveness, we train a single-stage anchor-free detector called Soft Anchor-Point Detector (SAPD). Experiments show that our concise SAPD pushes the envelope of speed/accuracy trade-off to a new level, outperforming recent state-of-the-art anchor-based and anchor-free, single-stage and multi-stage detectors. Without bells and whistles, our best model can achieve a single-model single-scale AP of 47.4% on COCO. Our fastest version can run up to 5x faster than other detectors with comparable accuracy.