90 days later: what remains of IBM’s three July 2026 quantum advantage claims?
As of October 30, 2026, none of the three quantum advantage claims IBM announced on July 30, 2026 has been peer-reviewed or adjudicated by a neutral referee. The flagship doped Clifford experiment was matched classically in 14 days — all 2,051 published samples, 37.3 minutes on 256 GPUs (arXiv 2608.13110) — and IBM revised the paper on September 2 with the challengers added as co-authors and a new 64-qubit instance (arXiv 2607.25941 v3). That new instance already has a classical reply filed on the Quantum Advantage Tracker (issue #249, consulted 2026-10-30). The other two claims stand unchallenged — and unreplicated. Rosetta Quantum scores the 90 days below, document by document.
What did IBM claim on July 30, 2026?
Three separate advantage claims, in three preprints, under one slogan. IBM's announcement — «quantum advantage through trusted quantum computation», with Jay Gambetta quoted saying «we are now firmly in the quantum advantage era» (IBM Quantum blog, 2026-07-30; PostQuantum fact-check, 2026-08-01) — bundled:
- Doped Clifford sampling (IBM + University of Chicago, arXiv 2607.25941, v1 posted 2026-07-28): a depth-70 circuit on 70 data qubits plus 27 syndrome ancillas (97 physical), 2,869 CZ gates, producing 2,051 accepted samples in 16.1 minutes of QPU time, with a certified state-fidelity lower bound of 0.284 at 95% confidence (PostQuantum, 2026-08-01). The one explicit, complexity-theoretic advantage claim of the trio.
- Error-mitigated Floquet dynamics (Qedma + RIKEN + BlueQubit, arXiv 2607.24937): Floquet Ising magnetization on Heron R3 at up to 74 qubits, where two state-of-the-art classical methods — sparse Pauli propagation runs of ≈10^12 Pauli strings on 12,888 Fugaku nodes — diverged from each other after ≈7 cycles at 51 qubits (PostQuantum, 2026-08-01). An empirical beyond-classical result, not a hardness proof.
- Operator Loschmidt echo (Algorithmiq, arXiv 2607.25998): a 56-qubit estimate presented as «the most credible description among the methods considered», with mitigated results agreeing within 0.03 across five experimental variants (PostQuantum, 2026-08-01).
The built-in novelty was the verification layer: spacetime codes that let the computation certify its own fidelity instead of leaning on an external proxy (IBM Quantum blog, 2026-07-30). Rosetta covered the opening round in August in our doped Clifford explainer; this post is the 90-day audit of all three.
What happened to the flagship doped Clifford claim?
It was matched classically in 14 days, and the paper was rebuilt around that fact. On August 13, Manabe, Gu and Pan posted a tensor-network contraction that computed exact output amplitudes for all 2,051 of IBM's published samples in 37.3 minutes on 256 NVIDIA H100 GPUs (32 nodes × 8 GPUs), with a largest intermediate tensor of 256 GiB — which the authors report is 256× smaller than IBM's own hardness estimate — and a log-XEB estimate of 0.35034 (95% CI [0.29763, 0.40305]) against IBM's certified lower bound of 0.284 (arXiv 2608.13110).
IBM's response will read, years from now, as the most interesting part of the episode. Versions v2 (September 1) and v3 (September 2) of the paper added Manabe, Gu and Pan to the author list (arXiv 2607.25941 version history), acknowledged the simulation — the v3 text cites the 37.3-minute run at ≈159 H100-GPU-hours and computes exact ideal probabilities for all 2,051 reported samples (Quantum Timeline milestone, added 2026-10-15) — and replaced the instance: the revised experiment runs 64 qubits on a ring (periodic boundary conditions), depth 73, 314 T gates, 76 physical qubits, with a certified fidelity lower bound of 0.349 at 95% confidence (arXiv 2607.25941 v3 abstract). The geometry change is explicit about its purpose: the project's tracker entry says the ring closes the loophole that the open-boundary 1D layout had left for tensor-network contraction (Quantum Advantage Tracker issue #245, consulted 2026-10-30).
Two things can be true at once, and here they are. The original 70-qubit advantage claim, as instantiated, did not survive — its published outputs were reproduced exactly on a GPU cluster in half an hour. And the response was the system working: challengers co-opted into the paper, the instance redesigned in public, the new target posted on an open tracker. That is adversarial collaboration, not damage control — but it is also not an adjudicated advantage.
Where does the revised 64-qubit experiment stand today?
It is an «active candidate» already taking classical fire. The Quantum Advantage Tracker — launched in November 2025 by IBM, Algorithmiq, the Flatiron Institute and BlueQubit (PostQuantum, 2026-08-01) — lists the revised circuit as entry ibm_boston_dcs_nq64_depth73_checks12 (64 logical / 76 physical qubits, depth 73, 314 T gates, quantum runtime 870 seconds) in the «active candidates» category (tracker issue #245, consulted 2026-10-30).
And round 2 has begun: a submission from Calcul Québec reports an exact rank-width contraction on CPU that evaluated probabilities for 1,389 of IBM's published Experiment-1 bitstrings at ≈382 seconds per probability — a makespan of 66,325 seconds (≈18.4 hours) on 8 nodes, ≈530,598 CPU-seconds serial — yielding a linear XEB of 0.390 (95% CI [0.325, 0.457]) against IBM's certified lower bound of 0.349 (tracker issue #249, consulted 2026-10-30). Read carefully: that is verification of IBM's samples (the XEB sits inside the certified band, consistent with IBM's bound), and simultaneously a demonstration that exact per-bitstring probabilities of the hardened instance are computable on CPUs — the same capability that preceded the fall of the first instance. What it is not, so far, is classical sampling of the full instance.
Meanwhile, IBM's July 30 marketing page carries no editor's note, no update and no mention of the classical simulation or the revision — we checked it today (IBM Quantum blog, consulted 2026-10-30). The arXiv record is 90 days ahead of the newsroom.
Are the Qedma and Algorithmiq claims still standing?
Yes — standing, unchallenged, and independently unreplicated. Neither the Qedma Floquet result nor the Algorithmiq Loschmidt echo was covered by the classical counterattack (Quantum Timeline, added 2026-10-15), and our searches today found no registered classical challenge to either.
The honest caveats from the papers themselves still apply. Qedma's late-time 51- and 74-qubit trajectories rely on QESEM-Extrapolated, a heuristic the authors flag as capable of introducing bias, and the 51-qubit cross-checks on Quantinuum H2 and Helios cover selected time points, not a full 74-qubit replication (PostQuantum, 2026-08-01). Algorithmiq's global rescaling is described by its own authors as a heuristic with no quantitative bound on its distance from the unknown ground truth (PostQuantum, 2026-08-01). «Unchallenged» is a statement about the scoreboard, not about hardness: nobody has spent 12,888 Fugaku nodes twice.
What does the 90-day scoreboard look like?
| Claim | Scale claimed (source) | What the 90 days did | Status 2026-10-30 | Source |
|---|---|---|---|---|
| Doped Clifford sampling (IBM + UChicago) | 70 data qubits, depth 70, fidelity ≥0.284 (arXiv 2607.25941 v1) | All 2,051 samples computed exactly in 37.3 min / 256 H100s (Aug 13); v3 adds challengers as co-authors + 64q ring instance, fidelity ≥0.349 (Sep 2) | Revised instance = «active candidate»; CPU reply filed | 2608.13110 · 2607.25941 |
| Floquet Ising dynamics (Qedma + RIKEN + BlueQubit) | 74 qubits; exact references to 35q (arXiv 2607.24937) | No classical challenge registered; 51q spot-checks on Quantinuum H2/Helios | Standing, unreplicated at 74q | PostQuantum |
| Operator Loschmidt echo (Algorithmiq) | 56 qubits; variants agree within 0.03 (arXiv 2607.25998) | No classical challenge registered; ground truth unknown by construction | Standing, unverifiable classically | PostQuantum |
| The «advantage era» announcement | — (IBM blog, 2026-07-30) | No editor's note or update after the classical match and the revision | Unchanged since Jul 30 (checked today) | IBM blog |
The pattern matches what Rosetta measured elsewhere this month: vendor claims move on claimant-controlled scoreboards (our D-Wave annealing audit), definitions soften under pressure (quantum utility vs advantage), and the only referees with teeth so far are external programs like DARPA's QBI (our explainer). The doped Clifford episode is, so far, the cleanest adversarial loop the field has produced — and it still has no referee verdict.
What we don't know
Whether the 64-qubit ring can be classically sampled. The CPU submission computes exact probabilities of published bitstrings; generating full samples at the certified fidelity is a different workload, and no one has posted one (tracker issue #249, consulted 2026-10-30). Whether the Qedma 74-qubit late-time dynamics are classically reproducible — an absence of registered challenges is not evidence of hardness, and the extrapolation heuristic leaves room for bias the authors themselves flag. What the Algorithmiq ground truth is — unknowable by construction, which is both the point of the method and the limit of the claim. Peer-review status: we found no journal publication of any of the three papers as of today. Whether the Quantum Advantage Tracker will ever issue formal adjudications, and under what rule; its categories today distinguish «active candidates» from retired ones, not verdicts. When any of these moves — a classical sample of the ring, a 74q replication, a journal decision, a tracker verdict — Rosetta will re-score this page with the same method.
Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim.