Leaders
Most accurate: Helios (Quantinuum) · not ranked last month
Error-corrected qubits: Helios (Quantinuum) · not ranked last month
Most qubits: Ytterbium atom array (Atom Computing) · not ranked last month
Fastest: Nighthawk r2 (IBM) · not ranked last month
Use it today: Nighthawk r2 (IBM) · not ranked last month
Most accurate: Helios (Quantinuum) · not ranked last month
Error-corrected qubits: Helios (Quantinuum) · not ranked last month
Most qubits: Ytterbium atom array (Atom Computing) · not ranked last month
Fastest: Nighthawk r2 (IBM) · not ranked last month
Use it today: Nighthawk r2 (IBM) · not ranked last month
Home › Rankings › Most accurate
Two-qubit gate fidelity
Most accurate
How often a two-qubit operation, the basic building block of a quantum program, works correctly. Higher is better.
Why it matters
Small differences matter. A useful program runs thousands of these operations, so errors pile up quickly. Accuracy is one of the clearest signs of how good the hardware really is.
Keep in mind
Companies often report their best pair of qubits, or an average across the chip. We only compare results QubitBenchmark has marked as measured the same way.
Two-qubit gate fidelity. Ranked on grade A and B evidence only. Data: QubitBenchmark release 2026-11.1.
| # | Move | Machine | Type | Result | Grade | Measured | Evidence |
|---|
| 1 | · not ranked last month |
HeliosQuantinuum |
Trapped ion |
99.92 % |
B |
Jun 2026 |
Evidence → |
| 2 | · not ranked last month |
H2Quantinuum |
Trapped ion |
99.84 % |
B |
2025 |
Evidence → |
| 3 | · not ranked last month |
WillowGoogle |
Superconducting |
99.67 % |
B |
Dec 2024 |
Evidence → |
| 4 | · not ranked last month |
Harvard–MIT rubidium atom arrayQuEra Computing |
Neutral atom |
99.54 % |
B |
Oct 2023 |
Evidence → |
Compared within one method, as grouped by QubitBenchmark: two-qubit-gate-fidelity@mean. This is the first ranking, so movement shows from next month. Equal results share a place.
Not ranked this month
Results for this measure that don't count here, and why. We show them so nothing is hidden.
| Machine | Grade | Why it isn't ranked | Evidence |
|---|
| Tempo IonQ | C | Grade C: announcement-level evidence is shown with a warning and never ranked. | Record → |
| Forte IonQ | B | Measured a different way, so not comparable here | Record → |
| Oxford Ionics two-qubit prototype IonQ | B | Measured a different way, so not comparable here | Record → |
| Ytterbium atom array Atom Computing | B | Measured a different way, so not comparable here | Record → |
| Cesium atom array Infleqtion | B | Measured a different way, so not comparable here | Record → |
| Ankaa-3 Rigetti Computing | C | Grade C: announcement-level evidence is shown with a warning and never ranked. | Record → |
| Cepheus-1-108Q Rigetti Computing | B | Measured a different way, so not comparable here | Record → |
| Garnet IQM Quantum Computers | B | Measured a different way, so not comparable here | Record → |
| Emerald IQM Quantum Computers | C | Grade C: announcement-level evidence is shown with a warning and never ranked. | Record → |
How others rank this
Other organisations' rankings, credited. They don't change ours.
Jülich Supercomputing Centre (Forschungszentrum Jülich) and Purdue University; arXiv preprint · 2025-05-28
Ranks: machines
Academic researchers ran one fixed optimisation workload, linear-ramp QAOA on weighted MaxCut, on 24 cloud quantum processors from six vendors. They measured how long each machine keeps a usable signal as the circuit gets wider (up to 156 qubits) and deeper (up to 10,000 layers). Three problem layouts were used: a 1D chain, the chip's native layout, and a fully connected graph. This is a one-off study, not a continuously updated leaderboard, though Metriq has adopted the same test.
- quantinuum_H2-1, fully connected graphs: only platform to pass at 50 and 56 qubits (largest certified run), r = 0.872 at 56 qubits, p = 3
- ibm_fez, 1D chain (100 qubits) and native layout, r = 0.808 (1D chain, fractional gates); r = 0.771 (native layout, p = 10)
- ibm_fez, deepest run attempted, p = 10,000 (about 990,000 ZZ gates), signal coherent to about p = 300
- ionq_forte_enterprise, fully connected graph pass at 28 qubits, r = 0.770
- Who's behind it
- Academic authors. Funding came from the German BMBF and a Purdue startup grant. IBM access was via the Quantum Collaborative (the paper states IBM does not endorse its views), Quantinuum access came through Oak Ridge with a capped budget, and some IBM results were community-contributed. Vendors did not author the paper.
- How much weight to give it
- Careful, independent academic study using one workload; it is a 2025 snapshot that predates newer machines, and the access budgets differed by vendor.
Free · Read it at Jülich Supercomputing Centre (Forschungszentrum Jülich) and Purdue University; arXiv preprint →
