Is quantum computing actually getting better?
Yes, quickly in some ways and slowly in others. This page answers the question with public data only: published experiments, real calibration records from machines in the cloud, and independent benchmarks. Every number links to where it came from, and every source is labelled by how much weight it can bear.
How to read the labels · Sources & method · Download the data (JSON)
In 1998, about 1 in 3.3 entangling operations failed. The best result so far (trapped ions, 2025) fails about 1 in 11,900.
Acharya et al. ran a distance-7 surface code on superconducting hardware below the error-correction threshold: adding qubits made the protected (logical) qubit better, not worse.
A typical IBM device's two-qubit gate failed about 1 in 63 times in 2019 and about 1 in 520 in 2026 (median across 8 devices).
That estimate assumes every qubit fails at most about 1 in 1,000 operations, and the earlier 2019 estimate needed 20,000,000. Better algorithms are closing the gap from one side, bigger and cleaner machines from the other.
How often do quantum operations fail?
A quantum computer works by entangling qubits, two at a time, and every one of those operations can go wrong. Each dot is a published record for one kind of hardware; the line follows the best result so far. Lower is better. Below the dashed line (roughly 1 failure in 100), error correction can start to win: extra qubits fix more errors than they add.
Is error correction working yet?
Useful quantum computers will spread each reliable "logical" qubit across many physical ones, so that errors can be caught and fixed. The code distance measures how many errors a code can catch: distance 7 can correct any 3. Bigger is better, but only once the hardware is below the threshold. Stars mark experiments reported as below threshold.
How big are the machines getting?
Qubit counts are the number everyone quotes, but on their own they say little: a thousand noisy qubits can do less than a hundred good ones. Read this chart together with the error charts. The biggest neutral-atom figures are atoms trapped in one array, not all yet usable in a single computation.
What about machines you can actually use?
Lab records are the best a team has ever managed. Cloud machines have to hold up across a whole chip, every day. IBM is the one vendor whose detailed calibration history is openly licensed, so here is every published snapshot of its public fleet: 202 snapshots from 68 devices, 2019–2026. Each dot is one device on one day; the line is the typical (median) device each year.
How far is it from breaking encryption?
Breaking RSA-2048, the encryption behind much of the internet, is the benchmark people worry about. Published estimates of what it takes have fallen sharply as algorithms improved, but they still assume machines far larger and cleaner than any that exist: every qubit failing at most about 1 in 1,000 operations. Compare that with the cloud fleet above.
Who is being tested by government?
DARPA's Quantum Benchmarking Initiative aims to test whether any approach can reach utility-scale quantum computing by 2033. In October 2026 it moved Atom Computing, Diraq, IBM, IonQ to Stage C, hands-on testing of their hardware, joining Microsoft and PsiQuantum, which reached the equivalent stage earlier. 11 companies had entered Stage B in November 2025.
Has quantum beaten classical computers yet?
Claims of "quantum advantage" are often answered within months by a cleverer classical method. The community-run Quantum Advantage Tracker, started by IBM, the Flatiron Institute, Algorithmiq and BlueQubit, with outside contributors, logs each claim and whether a classical method has since matched it. It's the right place to check a headline.
How do today's cloud machines compare?
Vendors measure their machines in their own ways. Metriq, run by the Unitary Foundation, runs the same open benchmarks across vendors. Provenova keeps every dated Metriq result alongside each device's calibration history, so you can see how a machine changes over time.
How to read the labels
Not all numbers deserve the same trust. Every figure on Provenova carries one of these.
- Published experiments
- A result reported in a paper (peer-reviewed or a preprint), usually a record on a few qubits in a lab.
- Device calibration
- The machine's own measurement of every qubit and gate on a given day, published by the vendor.
- Independent benchmark
- The same open test run on many vendors' machines by a third party (Metriq).
- Vendor claim
- A number from a company's own announcement or spec sheet, not independently checked here.
- Resource estimate
- A paper's calculation of what a future task would need. It depends on its stated assumptions.
- Programme announcement
- A decision announced by a funder or government programme, reported with its source.
Sources & method
- Published experiments: Awesome Quantum Computing Experiments by François-Marie Le Régent (MIT licence; paper), version f97919e of 2026-09-08. Spot a missing or wrong experiment? Corrections go to that project, and flow here.
- Device calibration: IBM calibration files published in Qiskit's fake backends (Apache-2.0), taken from their git history so each snapshot keeps its real date. Every snapshot links to the exact source file and its hash. How the corpus works.
- Independent benchmarks: every dated Metriq run (CC-BY-4.0), one snapshot per device per day.
- Context: resource estimates and programme decisions are hand-curated with links, as of 2026-10-20.
- Kept up to date automatically: new Metriq runs are collected daily, and new IBM calibration files and dataset versions weekly (last checked 2026-10-19). New data appears without anyone editing this page.
The data is committed to the open repository, so every change is reviewable. Found a problem? Open an issue. Running your own experiments? Record them with Provenova so your results carry the same receipts.