A vendor says 100 logical qubits. Here is what the labs have actually measured.
A qubit count is a claim. This is the field scored on real, sourced data: every named demonstration graded on whether its logical qubits can remember and whether they can compute, with what a lab measured kept strictly apart from what a vendor announced. The one number comparable across labs is the error-suppression factor, Lambda.
Below threshold, logical error falls as the code grows. Two labs have measured it.
Lambda is how much the logical error drops for each two steps of code distance. Above 1, the device is below threshold and adding qubits helps. The two points below are real measured devices, not a model.
The higher the Lambda, the steeper the drop, the more an extra layer of code buys. Below 1 it would mean error correction is making things worse.
Both lines share a starting point so the slopes compare directly. The slope is the measured Lambda factor: Willow 2.14, Zuchongzhi 1.40. The Willow line passes through its measured floor at distance 7, 0.143% logical error per cycle (Nature 2024). Vertical axis is logarithmic.
The measured field
Every named demonstration, scored on what it actually showed.
Two tiers. Memory is whether the logical qubit can remember, the five criteria. Compute is whether it can operate, the axes a fault-tolerant machine is judged on. Pick a filter to read one question across the whole field. Nothing here is a roadmap target.
Tap any of the nine questions above to read it across the whole measured field. Or read the full scoreboard below.
| Demonstration | Where | 1Beat HW | 2Λ | 3Cycles | 4No PS | 5Run | 6Gate | 7Magic | 8Decode | 9Clock |
|---|---|---|---|---|---|---|---|---|---|---|
| Google Willow Distance-7 surface code below threshold. Beats its best physical qubit by 2.4x. Google Quantum AI, Nature 2024 Cleanest beyond-breakeven on record. Real-time decoding demonstrated. A memory result, not a computation. | US Superconducting 2024 Λ 2.14 | |||||||||
| Zuchongzhi 3.2 107 qubits, distance-7 below threshold. First surface-code below-threshold memory outside Google. USTC, Phys. Rev. Lett., Dec 2025 All-microwave leakage suppression. Clean, non-post-selected below-threshold scaling, at a lower Lambda than Willow. | CN Superconducting 2025 Λ 1.4 | |||||||||
| Quantinuum + Microsoft Logical error rates 9.8x to 800x below physical, on H2 trapped ions. arXiv 2404.02280 Logical entangling operations shown. The 800x figure is reached under post-selection. [[12,2,4]] is 2 logical qubits, not 12. Slow logical clock (ions). | UK Trapped ion 2024 | |||||||||
| QuEra / Harvard Up to 48 logical qubits on 280 atoms, with logical algorithms. Bluvstein et al, Nature 2023 Largest logical-qubit count shown. Ran a non-Clifford (CCZ) logical gate. Many results rely on error detection and post-selection. | US Neutral atom 2023 | |||||||||
| Atom Computing + Microsoft 24 logical qubits in the [[4,2,2]] error-detecting code. Microsoft + Atom Computing, Nov 2024 On a 1,180-atom machine. [[4,2,2]] is distance-2: it detects errors by discarding shots, it does not correct them. | US Neutral atom 2024 | |||||||||
| Alice & Bob cat qubit Bias-protected cat qubit with bit-flip times exceeding 10 seconds. Reglade et al, Nature 2024 Bit-flip strongly suppressed by cat size, but phase-flip rises with it, so total error does not fall with scale. Needs an outer code. A building-block result. | FR Superconducting cat 2024 |
Japan (photonic continuous-variable) and Australia (silicon spin) have working architectures but no published below-threshold logical-memory result yet.
These are targets. They are on a different page on purpose.
A roadmap number is a plan, not a result. None of the figures below has been demonstrated. They never share a column with the measured field above.
A memory that remembers
Several labs have a logical qubit that holds information better than its physical parts, below threshold, measured honestly. Google and USTC have the cleanest. This is real progress, and it is the first four of the five criteria.
A machine that computes
A useful computer must also run logical gates, make magic states for the non-Clifford operations that matter, decode errors in real time, and run long enough to finish the job. Pieces have been shown in isolation: real-time decoding (Google), a non-Clifford logical gate (QuEra), magic-state distillation on logical qubits (Quantinuum, 2025). A logic set without magic states is classically simulable, so the easy gates alone are not the hard part. No public result has cleared the whole tier together, fault-tolerantly, at scale.
The gap a buyer is actually being sold
What is announced, beside what is measured.
The gap is years of engineering. The deadline on protecting data is not, because data with a long shelf life is being copied now to read later. The clock on migration runs separately from the clock on qubit counts.
Carry the real field state into the room
Five questions, and the measured answers as of 2026.
The questions chain. A number measured with too few cycles or heavy post-selection cannot be trusted, whatever else is true. And a memory result, however clean, is not yet a computer.
Carry these into the meeting
- Did the logical qubit outlive your best physical qubit, and what is the Lambda factor?
- Does the error rate fall as you add code distance, measured at more than one distance?
- How many rounds did you run, and were any runs post-selected away?
- Have you run a logical gate, and made a magic state, or only stored a qubit?
- Is this a measured result or a roadmap target, and what year is the target?
The honest 2026 picture is two sentences. A handful of labs across the US, China, and Europe have shown a logical qubit that remembers better than its hardware, and Lambda says which ones by how much. No one has shown the machine that computes on those qubits at useful scale, and the timeline for that is a range, not a date.
Put your own posture against this field: run the readiness scan →The five memory criteria are the work of Alice & Bob, published as "Defining the Logical Qubit" under CC BY-SA 4.0. We built the scorecard, the computation tier, and the measured-versus-announced reading on top of it. Source report at alice-bob.com. This adaptation is shared under the same CC BY-SA 4.0 licence.
Measured demonstrations, by primary source. Google Willow, distance-7 surface code, Lambda 2.14, 2.4 times its best physical qubit: Nature 2024. Zuchongzhi 3.2, distance-7 below threshold, Lambda 1.40: Phys. Rev. Lett., 2025. Quantinuum and Microsoft, logical error 9.8 to 800 times below physical: arXiv 2404.02280. QuEra and Harvard, up to 48 logical qubits with post-selection: Nature 2023. Atom Computing and Microsoft, 24 logical qubits, [[4,2,2]] code: 2024. Alice & Bob, cat-qubit bit-flip beyond 10 seconds: Nature 2024. Resource estimates for RSA-2048: Gidney and Ekera 2021 (20M qubits, 8 hours), Gidney 2025 (under 1M noisy qubits, under a week), Gouzien and Sangouard 2021 (177 days, 13,436 qubits). Roadmap targets are each labelled announced, not measured. Hardware figures cross-checked against our hardware roadmap registry.