The scoreboard on the labs racing to fault tolerance, dated and sourced. Our own prediction record lives on The Ledger, tracked by the same public rules as every other claim we follow, and the desk's own forecasts for this race are below. We keep score on ourselves so you can decide whether to trust us.
Qubit counts make the headlines. They are close to meaningless on their own. The number that decides whether a machine is in this race at all is Λ: how much the logical error rate shrinks each time you make the code bigger.
The rule, from the scaling law on the QEC page: raising the code distance by two (d → d+2, one step, which is what the word means everywhere below) divides the logical error rate by Λ. Raising an odd d by one buys nothing at all. Below threshold Λ > 1 and errors fall away geometrically as you add qubits. Above threshold Λ < 1 and the same extra qubits make the machine worse. Same hardware, same code, opposite direction, and the whole industry is sorted by which side of that line it sits on.
Logical error per round, relative to d=3: shorter is better. Logarithmic scale, no gridlines by design: the axis runs from 10⁻⁴ (floor) to 10² (full height), so a bar that is one sixth of the axis shorter than its neighbour is ten times better, and one that is half the axis shorter is a thousand times better. Read the number on the bar, not the picture. The count under each bar is 2d²−1 (data + ancilla for a rotated surface code); Willow's distance-7 code added 4 leakage-removal qubits, 101 in total.
Honest model: this is the standard below-threshold scaling form pL ∝ (p/pth)⌊(d+1)/2⌋, plotted as a ratio to d=3, which is why no fitted prefactor appears and none is needed. It is the shape of the curve, not a prediction of any specific chip's error rate. Real devices depart from it near threshold, and Λ itself drifts with calibration. Here Λ is set by the slider as the suppression per d → d+2 directly, and the scaling form above is what ties that to the ratio p/pth; in a real device the two agree only as well as the model does. Willow's Λ = 2.14 ± 0.02 is a measured suppression factor for increasing the code distance by two, obtained by fitting the log of the logical error rate against distance across the d = 3, 5, 7 series, not a single-step reading (the lone 5 → 7 ratio is ≈ 2.1) (arXiv 2408.13687, Nature 638, 920 (2025)).
Each lab's latest entry on the hardware table below, with its date and source there. The last column asks a different question: not what a lab has done, but what it has already promised on the record, with a date we will score it against whether or not it wants us to. Read straight across and the gap between the two is the story.
| Lab | Latest entry | What it was | Below threshold? | Promised, on the record |
|---|---|---|---|---|
| 2024-12 · Willow, 1 logical | One logical qubit, at distance 7, below threshold: the first whose error rate provably falls as you add physical qubits. | below threshold | no fault-tolerance claim tracked | |
| IBM | 2025-11 · Nighthawk, 120 physical | 120 qubits on a square lattice with 218 tunable couplers, over 20 percent more couplers than Heron. | not shown below threshold | on The Ledger → |
| Quantinuum | 2025-11 · Helios, 48 logical | 48 error-corrected logical qubits from 98 physical trapped-ion qubits, reported to perform better than the physical qubits underneath. | not shown below threshold | on The Ledger → |
| Microsoft | 2024-11 · neutral-atom system, 24 logical | 24 logical qubits entangled together on a neutral-atom machine, which Microsoft and Atom called a record for entangled logical qubits. | not shown below threshold | on The Ledger → |
| USTC | 2025-12 · Zuchongzhi 3.2, 1 logical | A distance-7 surface code with logical error suppression factor 1.40(6) per two steps of distance, and leakage suppressed with microwave control alone. | below threshold | no dated claim tracked |
| Harvard / QuEra | 2025-06 · 448-atom fault-tolerant architecture, 1 logical | A surface-code memory on up to 448 neutral atoms: distance 5 had 2.14(13)x lower error per round than distance 3 over a four-round circuit, using atom-loss detection and machine-learning decoding. | below threshold | no dated claim tracked |
No entry on the hardware table yet for IonQ or PsiQuantum.
Pre-registered on 2026-09-30: the wording, the probability and the resolver are fixed here before anything resolves, and each is scored with a Brier score on this page when its date passes, whichever way it goes. These are the desk's estimates, not anyone's claim, so they are not Ledger entries. The Ledger scores claims that other people made.
| Forecast | Our probability | Resolves by | Resolved from |
|---|---|---|---|
| IBM shows a Kookaburra module that both stores error-encoded information and processes it, the same test as the Ledger's Kookaburra claim. | 40% | 2026-12-31 | IBM announcement plus independent technical reporting or a paper |
| A primary source dated 2026-10-01 or later reports more than 48 logical qubits (Helios's count on the table below) in one encoded computation. | 55% | 2026-12-31 | The paper or lab announcement itself; we add it to the table below if it qualifies |
| The table below gains at least one more below-threshold entry (logical error falling with code distance, stated in a primary source) beyond the three flagged. | 60% | 2027-03-31 | arXiv or a peer-reviewed paper; the entry must state the suppression factor |
Why these three: each is decided by a document anyone can read, none depends on how a press release is worded, and the first ties to a claim the Ledger is already tracking so you can compare our odds with the record.
The scoreboard above lists each lab's latest entry. Below is the full record: a hand-curated set of milestones, each dated and sourced, on a log axis. Two things show up. Physical qubit counts climbed about twenty-fold from 2019 to 2023 — and then the leading superconducting chips stopped competing on raw count. Google's Willow (2024) holds under a tenth the qubits of 2023's largest chip. Neutral-atom arrays kept growing, to 6,100 atoms in 2025, but those are atoms held in traps, not qubits in one computation. The numbers that matter now are logical qubits, and there the counts are small: the one that is actually below threshold, the first that gets better as you grow it, is a single qubit.
“Qubit count” is not one number: this splits physical from logical, and even then a physical count spans fabricated sites, controlled qubits, and qubits good enough to compute with.Three entries here are below-threshold results, meaning a code whose logical error falls as it grows: Google Willow (2024-12); Harvard / MIT / QuEra 448-atom fault-tolerant architecture (2025-06); USTC Zuchongzhi 3.2 (2025-12). The other logical points are qubits encoded and operated, not proven to improve with scale. Curated by the desk from the sources in the table below; last updated2026-09-30 . No 2026 result has been added yet: none has met the bar of a new count stated in a primary source. The same data lives in hardware.json.
| Date | Org | Device | Kind | Qubits | What it was |
|---|---|---|---|---|---|
| 2019-10 | Sycamore | physical | 53 | The "quantum supremacy" chip: a sampling task in 200 seconds that Google argued would take a classical supercomputer millennia (a gap classical work later narrowed). | |
| 2021-11 | IBM | Eagle | physical | 127 | First IBM processor past 100 qubits; introduced the multi-level wiring that the later big chips scaled up. |
| 2022-11 | IBM | Osprey | physical | 433 | Roughly a 3.4x jump in a year, and the last of IBM's chips whose headline was its size. |
| 2023-10 | Atom Computing | second-generation array | physical | 1180 | Neutral atoms held in optical tweezers; the first system with more than 1,000 qubit sites. Sites in a trap array, not 1,180 simultaneously computing qubits. |
| 2023-12 | IBM | Condor | physical | 1121 | IBM's largest single chip, and deliberately its last of that kind: the roadmap released the same day pivots to smaller, lower-error modular processors. |
| 2023-12 | IBM | Heron | physical | 133 | Shipped alongside Condor as the actual direction: about a 3-5x lower error rate than Eagle, and the template for everything IBM has built since. |
| 2023-12 | QuEra / Harvard / MIT | neutral-atom processor | logical | 48 | 48 logical qubits from up to 280 physical atoms, with logical circuits run across them. Encoded and operated, not scaled below threshold. |
| 2024-04 | Microsoft / Quantinuum | H2 | logical | 4 | Four logical qubits with a reported ~800x lower error rate than the underlying physical qubits, running thousands of operations. |
| 2024-11 | Atom Computing / Microsoft | neutral-atom system | logical | 24 | 24 logical qubits entangled together on a neutral-atom machine, which Microsoft and Atom called a record for entangled logical qubits. Not the largest logical count (the 2023 entry above has 48), and still not a below-threshold result. |
| 2024-12 | Willow | physical | 105 | Under a tenth the size of 2023's largest chips, and about twice Sycamore's. The point was not the count: it was the first processor to show a surface code getting exponentially better as the code grew (below threshold), with the logical qubit outliving its best physical qubit. | |
| 2024-12 | Willow | logical | 1 | One logical qubit, at distance 7, below threshold: the first whose error rate provably falls as you add physical qubits. The number is 1 on purpose. | |
| 2025-06 | Harvard / MIT / QuEra | continuously reloaded atom array | physical | 3000 | Over 3,000 atoms kept in an array for more than two hours by reloading atoms as they are lost. The result is keeping a large array loaded and coherent, not a large computation. |
| 2025-06 | Harvard / MIT / QuEra | 448-atom fault-tolerant architecture | logical | 1 | A surface-code memory on up to 448 neutral atoms: distance 5 had 2.14(13)x lower error per round than distance 3 over a four-round circuit, using atom-loss detection and machine-learning decoding. Four rounds, not the long runs Willow reported. |
| 2025-09 | Caltech | tweezer array | physical | 6100 | 6,100 cesium atoms held in 12,000 laser tweezers, which Caltech called the largest qubit array ever assembled. Like the 2023 Atom Computing entry, these are atoms held in a trap array, not 6,100 qubits in one computation. |
| 2025-11 | IBM | Nighthawk | physical | 120 | 120 qubits on a square lattice with 218 tunable couplers, over 20 percent more couplers than Heron. IBM pitched it on connectivity and the number of gates a circuit can run, not on qubit count. |
| 2025-11 | Quantinuum | Helios | logical | 48 | 48 error-corrected logical qubits from 98 physical trapped-ion qubits, reported to perform better than the physical qubits underneath. Encoded and operated, not a demonstration that errors fall as the code grows. |
| 2025-12 | USTC | Zuchongzhi 3.2 | logical | 1 | A distance-7 surface code with logical error suppression factor 1.40(6) per two steps of distance, and leakage suppressed with microwave control alone. Below threshold, with a smaller margin than Willow's 2.14. |
Sources, in table order: Arute et al., Nature 574, 505 (2019), arXiv:1910.11333 · IBM Research announcements (15 Nov 2021, 9 Nov 2022, 4 Dec 2023) · Atom Computing announcement (24 Oct 2023) · Google Quantum AI, Nature 638, 920 (2025), arXiv:2408.13687 · Bluvstein et al., Nature 626, 58 (2024), arXiv:2312.03982 · Microsoft / Quantinuum announcement (3 Apr 2024) · Atom Computing / Microsoft announcement (19 Nov 2024) · Bluvstein et al., Nature 649, 39, arXiv:2506.20661 · He et al., Phys. Rev. Lett. 135, 260601 (2025).
Every prediction we make is falsifiable, dated, and scored when it resolves. No deleting misses. As of 2026-08-27 this table moved to The Ledger, the same mechanism generalised to track claims from anyone, not only from us. Our own predictions live there as ordinary entries (claimant = SymbiQ), scored by the exact same rules we hold everyone else to, published on our own scorecard rather than a separate table here. None made yet.
See every tracked claim, ours included, on The Ledger →
And the question underneath the whole race, when does a quantum computer first do something commercially or scientifically useful faster than classical, and on which hardware, is one nobody has settled.