Projection neurons (PNs) in the locust antennal lobe exhibit odor-specific dynamic responses. We studied a PN population, stimulated with five odorants and pulse durations between 0.3 and 10 s. Odor representations were characterized as time series of vectors of PN activity, constructed from the firing rates of all PNs in successive 50 ms time bins. Odor representations by the PN population can be described as trajectories in PN state space with three main phases: an on transient, lasting 1-2 s; a fixed point, stable for at least 8 s; and an off transient, lasting a few seconds as activity returns to baseline. Whereas all three phases are odor specific, optimal stimulus separation occurred during the transients rather than the fixed points. In addition, the PNs' own target neurons respond least when their PN-population input stabilized at a fixed point. Steady-state measures of activity thus seem inappropriate to understand the neural code in this system.