Results Explorer

How to read the results. The logical error rate is the fraction of shots in which any logical qubit failed over all syndrome-extraction rounds (by default as many rounds as the code distance), with a 95% Wilson confidence interval. The per-round rate assumes independent rounds: ε = (1 − (1 − 2P)^1/rounds)/2. Shots are sampled until 100 logical errors, 100,000 shots or one hour of decoding.

Status. measured at least 100 logical errors; low statistics stopped by the shot or time limit before 100 errors; upper bound no logical error seen, the 95% upper bound is given; does not fit the code needs more qubits than the device has; decoder not applicable e.g. matching on a code that is not matchable. "Qubits" counts the code's data and ancilla qubits, "circuit qubits" all qubits of the compiled circuit. Devices marked * are projections, not announced devices: a chain of as many Loon or Flamingo QPUs as the code needs, each linked to the next by four l-couplers. "Native layout" is a device built for the code, with a coupler for every pair of qubits its syndrome circuit entangles, so no SWAPs are needed; for bivariate bicycle codes this is IBM's planned layout, with its long-range couplers assumed to be as good as nearest-neighbour ones. "Native, noisy long-range" is the same layout with the two long-range couplers of each qubit as noisy and slow as IBM's remote couplers (3% error, 300 ns; 0.3% at tenfold lower noise). On devices with T1 and T2 (IBM, Infleqtion, Harvard/MIT), qubits decay for as long as they wait, for example while others are measured; IBM readout takes 1.56 µs, as on Heron today, and the tenfold-lower IBM model has threefold longer T1 and T2. Until 28 October 2026 waiting qubits did not decay and IBM readout took 70 ns; the results on these devices have been recomputed. Results with low statistics or that timed out are being refined with more shots in the background and are updated here as they finish.