We initiate the study of a new variant of the Maker-Breaker positional game, which we call multistage game. Given a hypergraph $\mathcal{H}=(\mathcal{X},\mathcal{F})$ and a bias $b \ge 1$, the $(1:b)$ multistage Maker-Breaker game on $\mathcal{H}$ is played in several stages as follows. Each stage is played as a usual $(1:b)$ Maker-Breaker game, until all the elements of the board get claimed by one of the players, with the first stage being played on $\mathcal{H}$. In every subsequent stage, the game is played on the board reduced to the elements that Maker claimed in the previous stage, and with the winning sets reduced to those fully contained in the new board. The game proceeds until no winning sets remain, and the goal of Maker is to prolong the duration of the game for as many stages as possible. In this paper we estimate the maximum duration of the $(1:b)$ multistage Maker-Breaker game, for biases $b$ subpolynomial in $n$, for some standard graph games played on the edge set of $K_n$: the connectivity game, the Hamilton cycle game, the non-$k$-colorability game, the pancyclicity game and the $H$-game. While the first three games exhibit a probabilistic intuition, it turns out that the last two games fail to do so.