Qubit counts and error rates by platform, 2019 to 2026

Qubit counts, two-qubit gate errors and logical qubits for five platforms, 2019 to 2026, each value traced to a paper or announcement.

6 min read

Key takeaways

  • Google’s Sycamore had 53 qubits and a 0.62 percent two-qubit gate error in 2019. The lowest peer-reviewed or preprint errors we found are about 0.08 percent, from a superconducting fluxonium pair (2023) and Quantinuum’s Helios (2025).
  • The largest arrays are neutral atoms: 6,100 trapped atoms in a 2025 paper and a vendor claim of 1,180 qubits in 2023. Neither source reports a two-qubit gate error for the whole array.
  • IonQ claims a gate fidelity above 99.99 percent. That is a vendor statement with no method given, so it is drawn hollow and kept apart from the papers.
  • Photonic machines have no comparable gate error. PsiQuantum reports a 99.22 percent fusion fidelity, which measures something else.

In 2019 the best-known quantum processor had 53 qubits and, in the Nature paper that announced it, an average two-qubit error of 0.62 percent. By late 2025 the table below holds a device with 98 qubits and an average two-qubit infidelity of 0.079 percent, and an atom array of 6,100. The two numbers did not move together. No row here pairs a count above 1,000 with a published gate error.

Three metrics, kept apart

Headlines mix three numbers, so we do not. A physical qubit count is how many qubits sit on one device or array. A two-qubit gate error is how often the basic entangling operation fails, taken as 100 percent minus the reported fidelity. A logical qubit count is how many error-protected qubits were shown, which depends on the code. Each has its own table, and a row appears only where a paper, a preprint or a vendor announcement gives the number.

Gate errors: superconducting and trapped ions lead

PlatformSystemDateTwo-qubit gate errorMetric as reportedSource typeSource
superconductingGoogle Sycamore2019-10-230.62%average two-qubit Pauli error e2, simultaneous operation (cross-entropy benchmarking)paperArute et al., Nature 574, 505, Fig. 2
superconductingMIT fluxonium pair2023-09-250.078%mean CZ gate fidelity 99.922 +/- 0.009 percent, averaged over about an hourpaperDing et al., Phys. Rev. X 13, 031035
superconductingUSTC Zuchongzhi 3.02025-03-030.38%reported two-qubit gate fidelity 99.62 percentpaperGao et al., Phys. Rev. Lett. 134, 090601
trapped ionQuantinuum race-track (H2)2023-12-180.184%average two-qubit gate infidelity 1.84(5)e-3 across the devicepaperMoses et al., Phys. Rev. X 13, 041052
trapped ionIonQ electronic qubit control2025-10-21under 0.01%fidelity 'exceeding 99.99 percent'; method not stated in the release, so 0.01 is an upper boundvendor claimIonQ press release, 21 Oct 2025
trapped ionQuantinuum Helios2025-11-070.079%average two-qubit gate infidelity 7.9(2)e-4 over all operational zonespaperRansford et al., arXiv:2511.05465 (preprint)
neutral atomHarvard/QuEra/MIT array2023-10-110.5%two-qubit entangling gate fidelity 99.5 percent, up to 60 atoms in parallelpaperEvered et al., Nature 622, 268
spinQuTech silicon processor2022-01-190.5%two-qubit gate fidelity above 99.5 percent, gate set tomographypaperXue et al., Nature 601, 343
spinRIKEN silicon processor2022-01-190.5%two-qubit gate fidelity 99.5 percentpaperNoiri et al., Nature 601, 338
photonicPsiQuantum silicon-photonics modules2024-04-26not a gate errortwo-qubit fusion fidelity 99.22 +/- 0.12 percent (a probabilistic measurement on two photons, not a gate); state preparation and measurement 99.98 percentpaperAlexander et al., arXiv:2404.17570
Error is 100 minus the reported two-qubit gate fidelity, or the reported infidelity. Benchmarks differ, so compare orders of magnitude, not the second digit. The IonQ figure is a vendor claim whose method is not stated.
Two scatter charts, 2019 to 2026. Left: two-qubit gate error by platform, from 0.62 percent (Sycamore, 2019) to about 0.08 percent for a fluxonium pair and Quantinuum Helios, with a vendor claim of 0.01 percent. Right: physical qubits per device, from 53 to 6,100 atoms.
Two-qubit gate error (left) and physical qubits per device (right). Hollow markers are vendor claims. The two metrics are never combined.

Two systems sit near 0.08 percent. An MIT fluxonium pair reached a mean controlled-Z fidelity of 99.922 percent, averaged over about an hour of operation with machine-tuned pulses. Quantinuum’s Helios, a 98-qubit trapped-ion machine, reports an average two-qubit infidelity of 0.079 percent over all its zones, in a preprint written by the vendor’s own team. A broad superconducting chip does worse: the 105-qubit Zuchongzhi 3.0 reports 99.62 percent, or 0.38 percent error.

Neutral atoms reached 99.5 percent on up to 60 atoms in parallel in 2023, and two groups reported the same for silicon spin qubits in 2022. Benchmarks differ: Sycamore’s figure is a Pauli error from cross-entropy benchmarking, the spin results come from gate set tomography, so small gaps between rows mean little.

Get the next one by email

One new article a week on radiation, particles and the instruments that measure them, with its sources. No spam, unsubscribe any time.

Counts: the biggest numbers are the loosest

PlatformSystemDatePhysical qubitsNoteSource typeSource
superconductingGoogle Sycamore2019-10-235353 qubits used of 54 on the chippaperArute et al., Nature 574, 505, doi:10.1038/s41586-019-1666-5
superconductingIBM Osprey2022-11-09433vendor announcement; no gate-error figure in the releasevendor claimIBM Newsroom, 9 Nov 2022
superconductingIBM Eagle2023-06-14127127 qubits of the Eagle processor used in the experimentpaperKim et al., Nature 618, 500, doi:10.1038/s41586-023-06096-3
superconductingGoogle Willow2024-12-09105qubit count from Google's announcementvendor claimGoogle blog, Willow chip
superconductingUSTC Zuchongzhi 3.02025-03-03105n/apaperGao et al., Phys. Rev. Lett. 134, 090601, doi:10.1103/PhysRevLett.134.090601
trapped ionQuantinuum race-track (H2)2023-12-1832initially operated with 32 qubitspaperMoses et al., Phys. Rev. X 13, 041052, doi:10.1103/PhysRevX.13.041052
trapped ionQuantinuum Helios2025-11-0798preprint by the vendor's own team, not peer reviewed at compilationpaperRansford et al., arXiv:2511.05465
neutral atomAtom Computing array2023-10-241,1801,225-site array populated with 1,180 qubits; vendor claimvendor claimAtom Computing press release, 24 Oct 2023
neutral atomCaltech tweezer array2025-09-246,100atoms trapped in about 12,000 sites; array size with coherent single-qubit control, not a full computerpaperManetsch et al., Nature 647, 60, doi:10.1038/s41586-025-09641-4
photonicXanadu Aurora2025-01-221212-qubit machine across 35 photonic chips; count from the vendor release, system described in the papervendor claimXanadu release (count); Nature 638, doi:10.1038/s41586-024-08406-9 (system)
spinIntel Tunnel Falls2024-12-251212-quantum-dot linear arrays on 300 mm waferspaperGeorge et al., Nano Lett., doi:10.1021/acs.nanolett.4c05205
Qubits on one device or array as the source reports them. Not every qubit is used in every circuit, and counts say nothing about quality.

IBM’s 433-qubit Osprey is the largest superconducting chip in the table, and its announcement gives no gate error. The two atom rows differ in kind: the Caltech paper traps about 6,100 atoms and shows coherent control, while Atom Computing’s 1,180 qubits is a company statement about a 1,225-site array. Xanadu’s photonic Aurora, at 12 qubits across 35 chips, is counted from the company’s release.

Logical qubits

Counting logical qubits is a different exercise. A 2023 neutral-atom processor ran up to 48 logical qubits, but with codes that only detect errors. Google’s Willow ran one distance-7 surface-code memory below the error threshold, with 0.143 percent logical error per cycle.

PlatformSystemDateLogical qubitsNoteSource typeSource
superconductingGoogle Willow2024-12-091one distance-7 surface-code memory on 101 qubits, 0.143 percent logical error per cycle, below thresholdpaperGoogle Quantum AI, Nature 638, 920, doi:10.1038/s41586-024-08449-y
neutral atomHarvard/QuEra/MIT logical processor2023-12-0648up to 48 logical qubits with [[8,3,2]] codes, which detect but do not correct errors; up to 280 physical qubitspaperBluvstein et al., Nature, doi:10.1038/s41586-023-06927-3
Logical qubits are counted separately from physical ones. The two rows use different codes and are not comparable.

Limits

This is a selection, not a census. We kept systems whose numbers trace to a paper or an announcement. Best values come from single devices or short runs, vendor claims are marked, and where a source gives no method we say so.

Cite this: ScienceShot, Qubit counts and error rates by platform, 2019 to 2026, November 2026, scienceshot.com, scienceshot.com/post/qubit-counts-error-rates-by-platform.

Sources: peer-reviewed papers, arXiv preprints and company releases from IBM, IonQ, Atom Computing, Xanadu and Google, each linked in the tables.

Cite this article

Citation style

Turner, A. (2026, November 17). Qubit counts and error rates by platform, 2019 to 2026. ScienceShot. https://scienceshot.com/post/qubit-counts-error-rates-by-platform

Turner, Alex. “Qubit counts and error rates by platform, 2019 to 2026.” ScienceShot, 17 Nov. 2026, scienceshot.com/post/qubit-counts-error-rates-by-platform.

Turner, Alex. “Qubit counts and error rates by platform, 2019 to 2026.” ScienceShot, November 17, 2026. https://scienceshot.com/post/qubit-counts-error-rates-by-platform.

@online{scienceshot-qubit-counts-error-rates-2026,
  author       = {Turner, Alex},
  title        = {{Qubit counts and error rates by platform, 2019 to 2026}},
  organization = {ScienceShot},
  date         = {2026-11-17},
  year         = {2026},
  url          = {https://scienceshot.com/post/qubit-counts-error-rates-by-platform}
}
Share

Get one article a week

One new article a week on radiation, particles and the instruments that measure them, with its sources. No spam, unsubscribe any time.

Alex Turner
Alex Turner

Alex Turner writes for the ScienceShot physics desk on radiation, particles and the instruments that measure them.

Articles: 176