Mitsui & Co. and Mitsubishi Electric published a joint benchmark on August 14, 2026, running the Quantum Fourier Transform on Quantinuum's 98-qubit Helios system two ways: as a raw physical circuit at increasing qubit counts, and encoded with the [[7,1,3]] Steane error-correcting code across up to 12 logical qubits. The two results, read together, are a genuinely useful side-by-side of what error correction buys you and what it currently costs.
The physical-qubit result: probability collapses as circuits grow
The team ran approximate QFT circuits (a truncated version of the full transform, a standard optimization since not every phase rotation contributes meaningfully to the result) scaling from 18 physical qubits up to Helios's full 98-qubit capacity. The target-state probability, how often the circuit produces the correct output, dropped from 0.976 at 18 qubits to 0.143 at 98 qubits. That's not a subtle decline. At full scale, the circuit produces the right answer well under a quarter of the time, purely from the accumulation of small gate errors and decoherence across a longer, wider circuit. This is the concrete version of a point this site has made in the abstract before: NISQ-era hardware runs into a real, measurable ceiling as circuits grow, not a theoretical one.
The logical-qubit result: much better, at a real cost
Encoding the same computation using the Steane code, which maps each logical qubit onto seven physical qubits, tells a different story. At 4 logical qubits (28 physical qubits), the target-state probability was 0.934, close to the small-scale physical result and far above the 98-qubit physical number. At 8 logical qubits, it dropped to 0.774, with only 31% of shots accepted under the error-detection scheme the team used. At 12 logical qubits, shot acceptance fell to 8%.
That acceptance rate is the part worth sitting with. The team compared two ways of handling detected errors: discarding runs where an error was flagged (error-detection post-selection) versus actively correcting the error and keeping the run (active error correction). Error detection produced higher target-state probabilities, but at the cost of throwing away most of the runs, exactly the tradeoff our error correction explainer describes: catching errors without disturbing the encoded data is only half the problem, deciding what to do once you've caught one is the other half, and every choice trades some combination of accuracy, speed, and yield.
The T-gate detail that cuts against the fault-tolerant narrative
One specific comparison in the benchmark runs counter to the usual story about fault-tolerant computing. The team compared applying a T-gate (a non-Clifford gate that most fault-tolerant schemes handle through expensive code-switching or magic-state distillation) two ways: the standard fault-tolerant code-switching approach, and a direct analog rotation that bypasses it. Under Helios's current noise levels, the direct analog rotation outperformed the fault-tolerant approach. That's a reminder that "fault-tolerant" techniques are a means to an end, lower logical error rates, not a guarantee, and at today's physical error rates a simpler, less-protected operation sometimes wins in practice.
What this adds to the Helios picture
Quantinuum's Q2 2026 earnings release included a vendor-reported claim of "near five-nines logical fidelity" on Helios using a novel error-correcting code, and Helios's original announcement centered on a 48-logical-qubit demonstration. This benchmark is a different, complementary kind of data point: an independent industrial user (not Quantinuum itself) running a specific, well-defined algorithm and publishing the probability and acceptance-rate numbers at each logical qubit count, using Quantinuum's own Guppy and pytket tooling. It doesn't confirm or contradict Quantinuum's fidelity claims directly, since it measures a different quantity, but it's a real external data point on the same hardware, with the messier, honest numbers that come from running an actual algorithm rather than a calibration benchmark.
What to watch next
Whether Mitsui and Mitsubishi Electric publish a peer-reviewed version of this benchmark, and whether the 12-logical-qubit shot-acceptance rate improves as Helios's physical error rates improve over future hardware generations. The acceptance-rate collapse from 31% at 8 logical qubits to 8% at 12 is the number that predicts whether logical-qubit computation is practical at a given scale, more directly than a headline fidelity figure.