logical qubit
A logical qubit is one qubit you can actually trust. Today's physical qubits are fragile: a stray bit of heat, a passing cosmic ray, or just the slow drift of time nudges them off course, and the quantum information leaks away in microseconds. The fix is the same idea you use when you read an important phone number back to someone twice: spread the information across many copies so that if a few get garbled, the rest still tell you the true answer. A logical qubit takes that idea and builds one reliable qubit out of many noisy physical ones, with extra qubits constantly watching for errors and correcting them on the fly.
Here is the catch, and it is why people keep quoting eye-watering qubit counts. The protection is not cheap. With a leading approach called the surface code, a single logical qubit can take hundreds to thousands of physical qubits, and that only works if each physical qubit's error rate is already below a threshold of roughly one percent. So a machine that needs, say, a hundred logical qubits to run a useful algorithm might need tens or hundreds of thousands of physical qubits underneath. A logical qubit is the unit a fault-tolerant algorithm is written for, but it is an expensive abstraction sitting on top of a lot of hardware.
This is also the honest dividing line between today and tomorrow. We are in the NISQ era: real machines have at most a handful of shaky logical qubits, or often none that are fully fault-tolerant. The famous algorithms that genuinely beat classical computers, like Shor's factoring, assume many clean logical qubits running long sequences of gates without drifting. Building enough logical qubits, cheaply enough, is the central engineering problem standing between the quantum computers we have and the ones the headlines describe.
A rough order-of-magnitude sketch with surface-code-style overhead; exact numbers depend on the code, the target error rate, and how noisy the hardware is.
A logical qubit is not faster than a physical one; it is more reliable, and that reliability is what long algorithms need to finish before noise wins.