IBM 70 Logical Qubits: Post-Selection Behind the Advantage Claim

What IBM and UChicago announced

On 30 July 2026, IBM and the University of Chicago said they had run a 70-logical-qubit circuit that leading classical simulation methods could not practically reproduce. They also said the run came with a way to estimate whether the quantum output could be trusted 1. The researchers argue that this pairing meets both halves of a proper quantum-advantage claim: classical intractability and verifiable fidelity 1.

The second half is the more interesting part. Producing output that a classical computer struggles to reproduce is easy to claim and hard to defend if nobody can show that noise didn't generate it. A fidelity bound that comes with the result is a real methodological step. Several roundups of 2026's biggest quantum advances now list the IBM work alongside Google's Willow error-correction milestone and Microsoft's logical-qubit work on Quantinuum hardware 5.

The numbers behind the headline

The technical details make the achievement more modest than "70 logical qubits" suggests. The circuit was 70 qubits wide and 70 layers deep, contained 468 non-Clifford T gates, and ran on just 97 physical qubits 2. That ratio is the first clue. Fault-tolerant error correction usually needs many physical qubits for each logical one. Here the team used spacetime error-detecting codes, which flag likely errors rather than fix them 2.

Detection helped. It cut the effective logical gate-error rate by roughly a factor of ten 2. But the cost appears in the post-selection rate of 5.90 × 10⁻⁴ 2. The system threw away every run in which an error was detected, and only about six in every 10,000 shots survived. The team collected 2,051 accepted samples in 16.1 minutes and reported a 95%-confidence lower bound on fidelity of 0.284 2. Broader coverage rounds the runtime to about 15 minutes 5. The small gap probably reflects how the figure was rounded rather than any real disagreement.

A fidelity floor of 0.284 is meaningful for a sampling benchmark, because it shows the outputs are not pure noise. It still means a large share of the signal is unaccounted for, even after aggressive filtering.

Why post-selection matters

This is where the coverage diverges. The more enthusiastic summaries present the result as a logical-qubit demonstration beyond classical reach and leave it there 5. The more detailed write-ups call it a narrow, device-dependent sampling demonstration. They say it is not evidence that quantum hardware has beaten classical machines on anything commercially useful 2. The longer explainers fall in between: they take the verifiability claim seriously while noting what a logical qubit actually is and is not 1.

My reading favors the cautious camp. Post-selection is a legitimate tool, but it does not scale. Discarding more than 99.9% of runs is workable for a sampling experiment, where you only need enough surviving samples to build a statistical picture. It does not work for long algorithms. As circuits grow, the chance of an error-free run falls exponentially, and so does the yield. True fault tolerance means correcting errors on the fly so a computation can keep going. This experiment does not do that. Calling the encoded units "logical qubits" is technically defensible, but the term invites comparison with the error-corrected logical qubits on industry roadmaps, and that comparison doesn't hold.

None of this makes the work trivial. Running hundreds of T gates under error detection, with a statistically grounded trust bound, is a real benchmark advance. It is a milestone on the road to fault tolerance, not the destination.

The cryptography question

Whenever logical-qubit counts jump, people ask whether "Q-Day" is coming sooner. Nothing here changes that picture. Breaking RSA or elliptic-curve cryptography would require deep, error-corrected computations sustained far longer than a post-selected sampling run can support.

The migration timeline is already set regardless. NIST's FIPS 203, 204, and 205 are final, and products can now claim compliance 3. FN-DSA is in draft as FIPS 206, and HQC was chosen in March 2025 as an additional key-encapsulation mechanism based on error-correcting codes 3. NIST guidance expects quantum-vulnerable algorithms to be deprecated by 2030 and disallowed by 2035. The NSA requires quantum-resistant cryptography for new national security system acquisitions starting in 2027 4. EU member states are expected to publish national PQC strategies by the end of 2026 4.

Those deadlines exist because migration takes years, not because of any one experiment. Results like IBM's are a reason to stay on schedule, not to panic or to relax.

The takeaway

IBM and UChicago have shown a verifiable, classically hard sampling run on error-detected logical qubits. That is an honest step forward with a carefully bounded claim. The weak points are the 0.06% survival rate and the reliance on detection instead of correction, and those are the gaps the next roadmap milestones will need to close.