Recent progress in code design has made it crucial to understand how quickly communication systems can approach their limits. To address this issue for the channel capacity C, we define the nonasymptotic capacity C/sub NA/(n, /spl epsi/) as the maximal rate of codebooks that achieve a probability /spl epsi/ of codeword error while using codewords of length n. We prove for the binary symmetric channel that C/sub NA/(n,/spl epsi/)=C-K(/spl epsi/)//spl radic/n+o(1//spl radic/n), where K(/spl epsi/) is available in closed form. We also describe similar results for the Gaussian channel. These results may lead to more efficient resource usage in practical communication systems.