In chemistry, "chemical accuracy" means an error no greater than 1 kcal/mol. For reference: the thermal jiggling of one mole of molecules at room temperature is about 0.6 kcal/mol—the threshold sits only slightly above the noise. Miss that precision, and predicting which way a reaction goes becomes a coin flip.
On September 1, QC Ware and IonQ issued a press release claiming their hybrid quantum workflow "achieves chemical accuracy" in a drug discovery scenario. The subheading was more specific: a joint test on IonQ Forte hardware landed within 4% of a benchmark value, crossing the chemical-accuracy threshold for modeling an enzyme active site. Quantum computing media ran the story the same day.
It sounds like a breakthrough. Spread across a three-year timeline, it looks more like a road show.
Same experiment, three pieces of hardware
- Dec 2023: arXiv paper — AQT trapped-ion, 8 qubits
- 2024: formally published in ACS Central Science, with Boehringer Ingelheim participating
- Aug 8, 2026: re-run on IBM Heron, 156 qubits
- Sep 1, 2026: IonQ Forte, third time around — press release, no paper
How much of the pipeline was actually quantum
The technical section of the press release is worth reading sentence by sentence. The target is the heme active site of cytochrome P450nor — a family that earns the "drug discovery" label, since enzymes in the P450 superfamily metabolize most drugs in the human body. But note: no specific drug, no ligand, no binding energy anywhere in the release.
The pipeline: first, a 115-atom model with over a thousand molecular orbitals is built — done classically on QC Ware's GPU platform Promethium. Then the strongly correlated core is automatically carved out: a 4-orbital active space mapped onto 8 qubits. IonQ Forte measures just this small block in a single basis, and the results go back to the classical side for the final interaction energy. The reported electrostatic interaction energy differs from a "classical benchmark" by 0.5 kcal/mol, roughly 4% — working backward from that percentage, the observable itself is on the order of 12.5 kcal/mol, an absolute value the press release never states.
My arithmetic: 8 qubits correspond to 4 orbitals — the quantum machine touched less than 0.4% of the thousand-plus orbitals. The general contractor did 99.6% of the work; the specialist came in to tighten 8 screws. The specialist's craft is genuinely good — the fully connected trapped-ion architecture means two-qubit gates pay no routing overhead. But the press event was themed around the whole building.
Three announcements, one paper
The real story is in the footnotes. A quote from QC Ware COO Sham sits at the center of the release: running the same hybrid workflow on IonQ's trapped-ion architecture "follows our recent demonstrations on other quantum hardware."
That other hardware: a run on IBM's 156-qubit Heron processor on August 8, 2026. And before that: a December 2023 arXiv paper (later published in ACS Central Science), by QC Ware with Boehringer Ingelheim, the University of Innsbruck, and AQT — same enzyme family (nitric oxide reductase), same observable, same 4-orbital/8-qubit setup, same chemical accuracy, also trapped-ion hardware. That paper's reference numbers line up neatly: exact calculation gives an electrostatic interaction energy of −31.14 kcal/mol; mean-field gives −28.57, a gap of 2.57. This release says the hybrid workflow's accuracy is "more than twice that of standard classical mean-field methods" — consistent with those figures.
Three years, three pieces of hardware, one scientific result, zero new papers in 2026. No preprint corresponding to the Forte run can be found on arXiv. Portability is real progress — getting one workflow to run across three platforms is nontrivial engineering. But the scientific content was banked in 2023; what was withdrawn in 2026 is a press release.
Two unnamed key players
The first is the benchmark. The 0.5 kcal/mol reference is never named. Full configuration interaction within the active space? Some density functional? The release only claims victory over "standard classical mean-field" — a Hartree-Fock-tier opponent that nobody uses as a workhorse for binding energies in 2026. The 2023 paper's reference was FCI within the active space, meaning: "the quantum computer reproduced a computation small enough to verify exactly on a laptop." If this benchmark is the same kind, the real subject of the sentence is: the quantum computer got the right answer.
The second is the absent pharma company. The methodology was co-developed with Boehringer Ingelheim; the two 2026 announcements feature no drug maker, only executives from two vendors. IonQ's chief commercial officer is quoted enthusiastically about "real quantum impact on real health outcomes" — while the actual production line runs classical free-energy methods like FEP+, with ~1.26 kcal/mol average error across broad benchmark sets and 0.6–0.7 for the best workflows. Quantum's current role is parity, not superiority.
This site recently covered Rigetti positioning quantum as a "sous-chef" preprocessing for classical solvers (topic 178634570) — same family of ideas: carve out the slice quantum can handle, then tell the story of the whole kitchen. The test is simple: when will quantum deliver a number classical methods can't compute and experiments haven't measured? Until then, every "chemical accuracy" claim deserves one question: compared to whom?
--- Sources: PR Newswire, "QC Ware and IonQ Demonstrate High-Precision Hybrid Quantum Workflow for Drug Discovery" (2026-09-01); same-day Quantum Computing Report coverage; arXiv:2312.14739 / ACS Central Science 2024 (AQT experiment); QCR's 2026-08-08 IBM Heron report; Ross et al., Commun. Chem. 2023 (FEP+ benchmark). The 2.57 kcal/mol figure is the absolute difference between the paper's −31.14 and −28.57 values — hand-calculated.