We discuss the efficiency of a heat engine operating in a nonequilibrium steady state maintained by two heat reservoirs. Within the general framework of linear irreversible thermodynamics we derive a universal upper bound on the efficiency of the engine operating at arbitrary fixed power. Furthermore, we show that a slight decrease of the power below its maximal value can lead to a significant gain in efficiency. The presented analysis yields the exact expression for this gain and the corresponding upper bound.PACS numbers: 05.70.Ln, 07.20.Pe The Carnot efficiency η C = 1 − T c /T h [1, 2] provides the upper bound on efficiency of heat engines working between two reservoirs at temperatures T h and T c , T h > T c . Though crucial from the theoretical point of view [3], practical applications of η C are rather limited, since the Carnot efficiency can be reached only when the heat engine operates reversibly. Reversible operation implies extremely long duration of the working cycle. As a result, when the engine efficiency reaches the upper bound η C , the output power is zero. Appealing universality of the upper bound η C , which depends solely on the two temperatures, and the needs of engineering solutions stimulated an intensive search for a more practical upper bound on the efficiency of heat engines operating at finite power. A promising candidate for which at least some universal properties can be derived was introduced about half century ago [4][5][6], it is the efficiency at maximum power η ⋆ .The upper bound on the efficiency at maximum power (EMP) in the linear response regime (linear in η C ) is equal to the famous Curzon-Ahlborn [7] formula η CA = 1 − T c /T h , which is to the linear order in η C equal to the half of the Carnot efficiency,The upper bound η ⋆ = η C /2 is achieved by a particular class of heat engines with strongly coupled thermodynamic fluxes. The assumption of strong coupling (see discussion below Eq. (5)) means that the heat flux is proportional to the flux, which generates work on the surrounding [9][10][11].In the present study we stay in the linear response regime (linear in η C ), however, we go beyond the regime of maximum power and study the engine efficiency at an arbitrary power P , 0 ≤ P ≤ P ⋆ (P ⋆ stands for the maximum power). One of the main messages is that the universal bounds on efficiency can be derived for an arbitrary P and not only at the point of maximum power which was considered in several recent studies [8][9][10][11][12][13][14], see however [15][16][17][18][19][20][21][22] for optimal regimes other than that with maximum power. To this end we introduce relative deviations from the regime of maximum power, the relative gain in efficiency δη and power δP :where −1 ≤ δP ≤ 0. Such normalization of the two principal engine characteristics allows us to derive several explicit results. One of them is that it is possible to provide a universal upper bound for the efficiency at an arbitrary power P . The bound depends explicitly on δP and it readsAt the maximum powe...