For a class $\mathcal{C}$ of binary linear codes, we write $θ_{\mathcal{C}}\colon (0,1) \to [0,\frac{1}{2}]$ for the maximum-likelihood decoding threshold function of $\mathcal{C}$, the function whose value at $R \in (0,1)$ is the largest bit-error rate $p$ that codes in $\mathcal{C}$ can tolerate with a negligible probability of maximum-likelihood decoding error across a binary symmetric channel. We show that, if $\mathcal{C}$ is the class of cycle codes of graphs, then $θ_{\mathcal{C}}(R) \le \frac{(1-\sqrt{R})^2}{2(1+R)}$ for each $R$, and show that equality holds only when $R$ is asymptotically achieved by cycle codes of regular graphs.