Quantum Computer Shootout

Everything here is runnable on your own circuit. Try it in the console

Everyone now wants to use a quantum computer, but which one should you choose, why, and for what? There are eight of them on this service, from seven manufacturers, across four different physical modalities.

This article describes them rather than ranking them. Each machine is listed with what it is, what the company that built it aims it at, and what it returned when we ran something on it. Nothing here is scored against anything else, and the reason becomes clear quickly. Three of the eight do not accept gate circuits at all, so no single benchmark exists that all eight could run.

The eight machines

Grouped by modality, because that is the property that decides what a machine can be asked to do. Qubit ceilings, shot windows and prices are read from the same registry that bills the job, so they are what you would actually be charged.

Superconducting

Rigetti Cepheus-1-108Q

108 qubits · 10 to 50,000 shots · $0.30 a task plus $0.000425 a shot

The widest gate processor on the service, built from superconducting transmons on a fixed lattice. Connectivity is nearest-neighbour, so a circuit needing distant qubits is compiled into chains of SWAP gates before it runs.

What the maker aims it at. Rigetti positions it for materials science, optimisation and quantum simulation, and for error-correction experiments that need enough physical qubits to encode a logical one.

Run here. A 3-qubit GHZ returned 40 of 50 shots in the two ideal states. It has also run satellite tasking, a portfolio QUBO and a Born machine.

qpu.rigetti

Superconducting

IQM Garnet

20 qubits · 1 to 20,000 shots · $0.30 a task plus $0.00145 a shot

A 20-qubit transmon device on a square lattice joined by tunable couplers. It is the smallest superconducting register here, and in our runs it was the quickest to return a result.

What the maker aims it at. IQM positions its processors for work in finance, energy, pharmaceuticals and logistics.

Run here. A 3-qubit GHZ returned 48 of 50 shots in the ideal states, and the H2 molecule returned a ZZ expectation of -0.9326 over 4,096 shots against an exact -1.0000.

qpu.iqm.garnet

Superconducting

IQM Emerald

54 qubits · 1 to 20,000 shots · $0.30 a task plus $0.0016 a shot

A 54-qubit transmon device with full square-lattice connectivity. It is built on the architecture Garnet uses, and IQM reports longer coherence and higher gate fidelities on it than on the smaller machine.

What the maker aims it at. IQM cites molecular simulation for photodynamic cancer therapy and 3D fluid-dynamics simulation among the work customers have run on it.

Run here. A 3-qubit GHZ returned 48 of 50 shots in the ideal states. It has also run satellite tasking and a portfolio QUBO.

qpu.iqm.emerald

Trapped ion

IonQ Forte Enterprise 1

36 qubits · 100 to 5,000 shots · $0.30 a task plus $0.08 a shot

Trapped ions with all-to-all connectivity, so any qubit can be entangled with any other directly. A circuit needing distant interactions pays no routing cost, which is where this modality differs most from a fixed lattice.

What the maker aims it at. IonQ positions it for drug discovery, optimisation and simulation, and cites enterprise work including telecommunications network optimisation.

Run here. A 3-qubit GHZ returned 95 of 100 shots in the ideal states. Its shot floor is 100, so it is the one device here that refuses a 50-shot task.

qpu.ionq

Trapped ion

AQT IBEX Q1

12 qubits · 1 to 2,000 shots · $0.30 a task plus $0.0235 a shot

Twelve fully connected calcium ions housed in two standard 19-inch racks, at room temperature and under 2 kW. It is the device here that installs alongside conventional computing equipment rather than around a cryostat.

What the maker aims it at. AQT positions it for proof-of-concept work in chemistry, portfolio optimisation, risk analysis and quantum security.

Run here. A 3-qubit GHZ returned 20 of 20 shots in the ideal states, and the H2 molecule returned a ZZ expectation of -0.9200 over 100 shots against an exact -1.0000.

qpu.aqt.ibex

Neutral atom, analog

QuEra Aquila

256 qubits · 1 to 1,000 shots · $0.30 a task plus $0.01 a shot

256 rubidium atoms held in optical tweezers and driven as a single analog system. A program is an atom layout and a pulse sequence rather than a gate circuit, so a gate-based benchmark does not run on it at all.

What the maker aims it at. QuEra describes Aquila as an analog Hamiltonian simulator, aimed at combinatorial optimisation and many-body physics.

Run here. Maximum independent set on a 3x3 register found the optimum, with 883 of 947 shots obeying the blockade constraint and 574 of them optimal.

qpu.quera.aquila

Neutral atom, analog

Pasqal FRESNEL

100 qubits · 1 to 100 shots · machine time at EUR 500 an hour, about EUR 0.56 a shot

The same analog model as Aquila on a second machine, so a Pulser sequence moves between the two by changing one argument. It is billed as machine time rather than per shot, which changes how a run is budgeted.

What the maker aims it at. Pasqal positions FRESNEL for analog quantum simulation and optimisation.

Run here. A 4-atom sequence at 20 shots certified at a hardware fidelity of 0.9398 against an exact Schrodinger evolution.

qpu.pasqal.fresnel

Photonic

Quandela Belenos

12 qubits · 1 to 800,000 shots · EUR 0.30 a job, whatever the sample count

A linear-optics machine where a program is an optical mesh plus an input photon state. It runs at room temperature in a standard rack, and it is billed per job rather than per shot, so sample count is close to free.

What the maker aims it at. Quandela positions Belenos for quantum machine learning, image sorting and generation and accelerated AI calculus, and has reported a validated low-latency path between its photonic processors and NVIDIA AI infrastructure.

Run here. Three runs returned Hong-Ou-Mandel interference visibilities of 95.4%, 97.0% and 97.7%, for about a euro each.

qpu.quandela.belenos

The use cases above are each manufacturer's own aim for its machine, quoted rather than endorsed. Where we have taken a use case end to end ourselves and published what came back, the sector benchmarks are collected on applications, covering chemistry, finance, logistics, scheduling, space, traffic, manufacturing and AI, each with the classical baseline it was measured against and the instance seeded so you can rebuild it.

Availability is the figure that changes hour to hour rather than month to month. Several of these devices accept work only during scheduled windows, so live engine status is worth checking before a submission. It reports which machines are open right now and how many tasks are queued ahead of you.

Every machine, with one measured run

Each row below is a real job on the service. They are deliberately different problems at different shot counts, because the machines do not share a common workload, and they are presented without commentary for the same reason. A table of eight incomparable runs sorted into an order would imply a ranking the data cannot support.

Every machine, one measured run each

Each row is a real job on the service at the price any customer pays. The rows are different problems at different shot counts, so they rank nothing against one another.

Maker	Machine	Modality	What it ran	Result	Shots	Cost
	Rigetti Cepheus-1-108Q	Superconducting	Satellite tasking, 14 requests	27.0176 against an exhaustive 32.5128	500	$0.5125
	IQM Garnet	Superconducting	H2 ground state, ZZ expectation	-0.9326 against an exact -1.0000	4,096	$6.239
	IQM Emerald	Superconducting	Satellite tasking, 14 requests	26.9421 against an exhaustive 32.5128	500	$1.100
	IonQ Forte Enterprise 1	Trapped ion	3-qubit GHZ state	95 of 100 shots in the two ideal states	100	$8.300
	AQT IBEX Q1	Trapped ion	H2 ground state, ZZ expectation	-0.9200 against an exact -1.0000	100	$2.650
	QuEra Aquila	Neutral atom, analog	Maximum independent set, 9 atoms	found the optimum; 883 of 947 shots valid	1,000	$10.300
	Pasqal FRESNEL	Neutral atom, analog	4-atom analog sequence	certified fidelity 0.9398	20	EUR 500/hour
	Quandela Belenos	Photonic	Hong-Ou-Mandel interference	visibility 95.4%, 97.0% and 97.7%	3 runs	~EUR 1 each

On cost, every gate task carries a flat $0.30 fee and then the device's own per-shot rate, which runs from $0.000425 on Rigetti Cepheus-1 to $0.08 on IonQ Forte Enterprise 1. The same 1,000-shot circuit is $0.725 on the first and $80.30 on the second. Two machines sit outside that model entirely. Quandela Belenos charges EUR 0.30 a job however many samples you take, and Pasqal FRESNEL bills machine time at EUR 500 an hour. What any of that comes to for a particular workload is worth pricing before you submit.

The four modalities, and what follows from each

The differences that matter between these machines are mostly structural rather than a question of quality, and they follow from the physics each one uses.

Superconducting processors run fast gates on a fixed lattice, which means two qubits interact directly only if they are neighbours. A circuit that needs distant qubits to interact is compiled into chains of SWAP gates first, and that expansion is the cost of the modality. Rigetti, IQM Garnet and IQM Emerald are the three here.

Trapped ion processors hold ions in an electromagnetic trap with all-to-all connectivity, so any pair can be entangled directly and the routing expansion above does not occur. Gates are slower and coherence is longer. IonQ Forte Enterprise 1 and AQT IBEX Q1 are the two here, and AQT's is the one machine in the fleet that runs at room temperature in standard racks.

Neutral atom machines in analog mode are not gate computers at all. Atoms are placed with optical tweezers and the whole register is driven by a pulse sequence, so the program is a layout and a Hamiltonian rather than a circuit. The Rydberg blockade forbids nearby atoms from both being excited, which makes independent-set problems a native fit rather than something encoded. QuEra Aquila and Pasqal FRESNEL are the two here, and a sequence written for one runs on the other by changing one argument.

Photonic machines compute with interference between single photons through an optical mesh. There is no cryostat, the device runs at room temperature, and because Belenos is billed per job rather than per shot, taking more samples is close to free. Quandela positions this modality at quantum machine learning and AI workloads specifically, and has reported a validated low-latency path between its processors and NVIDIA AI infrastructure.

Where a problem lands is usually decided by this list rather than by qubit counts. An independent-set or scheduling problem maps onto the neutral-atom devices directly; a chemistry ansatz or an optimisation circuit needs gates; a circuit with long-range interactions pays less on trapped ions; and a machine-learning sampling workload is the case the photonic device is aimed at. Each of those has been run and written up under applications if you would rather read the measured version than the argument.

One circuit, on the five machines that take gate circuits

A three-qubit GHZ state, built from one Hadamard and two CNOTs, then measured. It is the simplest circuit whose output is unambiguous. A perfect machine returns only 000 and 111, split evenly, so the fraction of shots landing in those two states is a direct if crude fidelity measure.

Shot counts differ because the devices impose different floors and ceilings, and the counts are small deliberately. This is what a first exploratory run costs rather than a benchmark campaign, and the section after the table says exactly what that constrains.

device                modality          qubits  shots   in GHZ   fidelity   queue+run   cost
Rigetti Cepheus       superconducting      108     50    40/50      80%      0.6 min  $0.3213
IQM Garnet            superconducting       20     50    48/50      96%      0.4 min  $0.3725
IQM Emerald           superconducting       54     50    48/50      96%     59.7 min  $0.3800
AQT IBEX Q1           trapped ion           12     20    20/20     100%     79.9 min  $0.7700
IonQ Forte Ent 1      trapped ion           36    100   95/100      95%      5.2 min  $8.3000

Cost is the provider's list price passed through without markup. Queue+run is wall-clock time from submission to result, so it includes waiting for the device's availability window rather than only the execution.

What these numbers will not support

The limits of the table above, stated plainly, because they are larger than the table looks.

Rigetti's 80% against IQM's 96% at 50 shots each is a two-proportion z of 2.46, so that one gap is statistically real. Nothing else here is.

The 95% confidence interval on a 50-shot run at 96% is roughly plus or minus 5 points, and AQT's 20 for 20 has a lower bound near 83% despite reading as a perfect score. IQM Garnet, IQM Emerald, AQT and IonQ are not distinguishable from one another on this data, and any article that ranked them from these figures would be inventing precision it does not have.

Price and measured fidelity did not track each other in this sample. The most expensive run was not the most accurate one. That is an observation about five small runs on one shallow circuit rather than a finding about the machines, and it is stated here only to discourage reading the cost column as a quality column.

One circuit is also not a benchmark. A three-qubit GHZ uses two entangling gates and finishes almost immediately, so it barely probes decoherence and says nothing about connectivity, which is where the trapped-ion machines differ most. A deeper circuit needing gates between distant qubits would compile into long SWAP chains on the superconducting devices and run natively on the ion traps, and the ordering above could invert entirely.

Queue times are a snapshot of one afternoon, not a service level.

Availability windows

The queue-and-run column is the one that changes how you work. IQM Garnet returned in 0.4 minutes. AQT IBEX returned in 79.9 minutes, and a separate single-shot task on the same device took 178.6 minutes.

For interactive work that distinction matters more than the price, which is why live engine status is published rather than asking anyone to trust a number measured once.

The same devices on a real workload

The GHZ circuit above measures the machines. This is several of them doing a job, satellite observation tasking at 14 requests, the same problem at 500 shots on every engine that takes it.

Run on our engines

Satellite observation tasking at 14 requests, seed 20260902, whose exact optimum is value 32.5128. Submitted to each kind of compute we offer, on 16 September 2026 at 500 shots. Every figure below is a real job on the service, priced as any customer would be priced.

Device	Engine	Kind	Qubits	Result	Cost
	mps.quimb.cpu	CPU	14	25.1546, gap 7.36 certificate	$0.0001
	exact.cpu	CPU	14	20.0569, gap 12.46 certificate	$0.0001
	exact.gpu	GPU	14	25.1546, gap 7.36 certificate	$0.0001
	qpu.iqm.garnet	QPU	14	26.8778, gap 5.64 certificate	$1.025
	qpu.rigetti	QPU	14	27.0176, gap 5.49 * certificate	$0.5125
	qpu.iqm.emerald	QPU	14	26.9421, gap 5.57 certificate•best outcome	$1.100
	neural.tpu	TPU	—	the tasking QUBO is diagonal, which is not the shape a neural ansatz is for	—

* The Rigetti row is a separate sample of ours on this same instance, with the QAOA angles re-optimised for it. The steps are in the docs.

A note on the hardware certificates: they state Hellinger fidelity against the exact distribution. For an optimisation circuit that distribution is spread across many outcomes rather than concentrated on one, so the figure is low by construction and is not a measure of whether the device found a good answer. The result column above is.

The same problem is yours to run: every instance here is seeded, so it rebuilds exactly. Open the console and a cost estimate is free before anything executes.

The full method, the classical baseline and what these figures do and do not show are on the space and satellites benchmark. The two IQM devices place differently on other problems in the same set, which is what a single unoptimised draw looks like.

Reproducing this

Every run above is a job on this service and can be repeated with a different engine name:

import qsim_sdk
c = qsim_sdk.Client(token="...")

for engine in ("qpu.rigetti", "qpu.iqm.garnet", "qpu.iqm.emerald",
               "qpu.aqt.ibex", "qpu.ionq"):
    est = c.estimate(ghz_circuit, shots=50, engine=engine)
    print(engine, est["predicted_cost_usd"])

The estimate is free and returns the exact charge before anything is submitted, which on hardware is worth checking. The same 50-shot request is $0.32 on one device and would be refused outright on IonQ, whose floor is 100 shots. The engine reference lists every device's bounds, and why quantum circuits get rejected covers the refusals you will meet first.

For the analog and photonic machines the unit of work is different, and each has its own write-up. See why neutral-atom machines are not gate-based, maximum independent set on neutral atoms and what the photonic device produced on a real run.

Pricing every figure below against your own workload: open the per shot cost calculator.