We study how to achieve the ultimate power in the simplest, yet non-trivial, model of a thermal machine, namely a two-level quantum system coupled to two thermal baths. Without making any prior assumption on the protocol, via optimal control we show that, regardless of the microscopic details and of the operating mode of the thermal machine, the maximum power is universally achieved by a fast Otto-cycle like structure in which the controls are rapidly switched between two extremal values. A closed formula for the maximum power is derived, and finite-speed effects are discussed. We also analyze the associated efficiency at maximum power showing that, contrary to universal results derived in the slow-driving regime, it can approach Carnot's efficiency, no other universal bounds being allowed.'fast-driving' regime, i.e. when the driving frequency is faster than the typical dissipation rate induced by the baths, which has received little attention in literature [50][51][52].By applying our optimal protocol to heat engines and refrigerators, we find new theoretical bounds on the efficiency at maximum power (EMP). Many upper limits to the EMP, strictly smaller than Carnot's efficiency, have been derived in literature, such as the Curzon-Ahlborn and Schmiedl-Seifert efficiencies. The Curzon-Ahlborn efficiency emerges in various specific models [53][54][55], and it has been derived by general arguments from linear irreversible thermodynamics [56]. The Schmiedl-Seifert efficiency has been proven to be universal in cyclic Brownian heat engines [57] and for any driven system operating in the slow-driving regime [24]. By studying the efficiency of our system at the ultimate power, i.e. in the fast-driving regime, we prove that there is no fundamental upper bound to the EMP. Indeed, we show that the Carnot efficiency is reachable at maximum power through a suitable engineering of the bath couplings. This is our second main results, illustrated in figures 2(b), (c) and figure 3. In view of experimental implementations, we assess the impact of finite-time effects on our optimal protocol, finding that the maximum power does not decrease much if the external driving is not much slower than the typical dissipation rate induced by the baths [58,59]. Furthermore, we apply our optimal protocol to two experimentally accessible models, namely photonic baths coupled to a qubit [22, 60-63] and electronic leads coupled to a quantum dot [21,23,58,59,64,65].C , and Γ in (c) denotes * G( ) H . 4 In principle, one can consider a broader family of controls including the possibility of rotating the Hamiltonian eigenvectors; however there is evidence that such an additional freedom does not help in two-level systems [45, 46]. ≔ [ ] ( ) J J t p t J J t p t