Luxonis OAK 4 cameras run up to 52 TOPS of AI on board from $749

The cheapest camera in Luxonis's OAK 4 line lists at $749 and carries 52 TOPS of AI inference on board, with no industrial PC, no GPU box and no video stream to the cloud required to make it work. The Denver company announced the four-device line and its companion cloud platform, Hub, in a Business Wire release on December 11, 2025. Robotics 24/7 and Vision Systems Design reported the specifications the following week.

The figure Luxonis leads with is the 40X. The company says 52 TOPS is 40 times the inference capacity of the previous OAK generation, according to both Business Wire and Robotics 24/7. Neither report states the prior generation's TOPS figure, so the multiplier is the company's framing rather than a benchmark a buyer can reconstruct from the announcement alone.

The direction is clear enough regardless. A camera that runs several models at once at high frame rates, including Luxonis's own Neural Stereo Depth models, changes the bill of materials for an inspection cell or an autonomous mobile robot. The host computer that used to sit next to the camera becomes optional.

What sits inside the IP67 housing

Every OAK 4 device runs on a Qualcomm QCS8550 processor that combines an image signal processor, a six-core CPU with GPU and a neural processing unit, all on Yocto Linux, Vision Systems Design reported. Depending on the model, the sensor stack includes a stereo depth pair, a 48-megapixel RGB camera built on Sony's IMX586 rolling-shutter sensor, an IR dot projector, an IR illumination LED, a nine-axis IMU and a microphone. A wide field of view option is aimed at navigation.

All of it lives in an IP67-rated aluminum chassis that Business Wire's release describes as engineered for shock, vibration and harsh environments. Power comes over USB or PoE, and an M8 connector handles IO. That connector detail matters more than it looks: it is the difference between a camera that bolts into an existing control panel and one that needs an adapter.

The feature worth reading twice is Dynamic Calibration, a patent-pending function that Business Wire says keeps stereo depth performance precise as environmental conditions change in the field. Stereo depth is only as good as the alignment between its two cameras. For a plant where cameras ride on forklifts or sit near presses, whether that alignment holds between maintenance windows would decide whether measurement data can be trusted across a shift.

Luxonis CEO Bradley Dillon framed the design as a response to what field deployments asked for: more edge compute throughput, more sensing modalities, higher-resolution imaging pipelines and tougher hardware, according to the Business Wire announcement. The sources describe no specific shortcomings in earlier OAK units, only the requests that shaped this one.

Inference stays local, management goes to the cloud

The Business Wire release states the architecture plainly: AI workloads run fully on the device and can be managed remotely with no additional hardware, which Luxonis says cuts server compute costs and removes the need to stream video to the cloud. Robotics 24/7 repeated the framing, describing the units as standalone spatial-AI sensors that need no external host system or cloud compute.

That is a direct answer to the question of which processing belongs at the edge for this class of device. Inference happens on the camera. Deployment, updates and monitoring happen in the cloud. Video, in this design, never has to leave the building.

For operators running sites with thin bandwidth, or with policies against camera footage leaving the premises, an architecture that runs AI workloads on the device and, according to Business Wire, eliminates the need to stream video to the cloud is a different conversation with IT than one built on streaming. It is also a different conversation about the camera itself. Each OAK 4 is a Linux computer, which puts it in the category of networked devices IT manages, not only the category of sensors the controls team wires up.

Hub, Snaps and the drift problem

Hub is the cloud half of the launch. Business Wire describes it as the platform that turns Luxonis hardware into a complete computer vision system, handling deployment, over-the-air updates and device monitoring. Vision Systems Design characterized it as built for plug-and-play deployment by users at every level of expertise, bringing systems, models and devices into one place.

Luxonis calls model drift a primary reason edge AI fails at scale, and Hub's answer is a feature called Snaps, which collects data at the edge and uses it to improve models, per the Business Wire release. The company also claims a workflow that once required specialized engineering teams can be completed in about five minutes with limited technical experience. That is an announced capability, not an observed result, and the published reporting includes no field data on how well Snaps holds accuracy over time.

Snaps is the part to put under a pilot. The sharper questions a buyer can now ask are how much data leaves the site through it, who labels that data and what the retraining cadence looks like. The sources do not answer those, and they are the questions that decide whether the drift story holds on a real line.

There is an install base behind the software. Luxonis reports more than 4.5 million downloads of its SDK, which Business Wire says wraps sensor control, stereo depth, neural inference and hardware acceleration into composable APIs. For a team that has to staff a vision project, that figure suggests the developer pool familiar with the toolchain is not small.

Four models, four prices

According to Business Wire, the OAK 4 S carries a $749 retail price, the OAK 4 D sells for $849, and the OAK 4 D Pro is priced at $949. Priced from $899, the OAK 4 CS was offered through Early Access at launch, with availability dates aimed at early 2026, while the other three could be bought immediately. Hub is sold on tiered plans that include a free developer option requiring no card.

Retail prices run from $749 for the OAK 4 S to $949 for the OAK 4 D Pro, according to Business Wire. What the announcement does not include is any comparison to the cost of a conventional camera paired with an industrial PC or GPU host. That total cost of ownership case is one a procurement lead has to build: the camera price against the host computer, the accelerator card and the cabling it displaces, plus whichever Hub tier a fleet of that size needs.

Qualcomm's stake in the launch is explicit. Anshuman Saxena, vice president and general manager for ADAS and robotics at Qualcomm Technologies, framed the OAK 4 and Hub pairing as an answer to one of the harder problems in scaling AI, keeping accuracy and adaptability intact in real-world environments, according to Business Wire. He positioned it for both robotics and advanced driver-assistance systems.

Where the line is pointed

Vision Systems Design positioned OAK 4 for 3D vision work in particular: precision measurement, gauging, industrial applications and inspection across varied environments. Business Wire and Robotics 24/7 describe the broader OAK install base as spanning industrial equipment, smart cities, manufacturing, retail and logistics, with applications running from AMRs to IoT sensors.

One point of difference between the reports is worth noting. Vision Systems Design describes the platform as reaching up to 52 TOPS, while Business Wire and Robotics 24/7 state 52 TOPS as the figure for the lineup. Since all four devices share the same QCS8550 processor, the gap reads as a difference in phrasing rather than in silicon, but a buyer comparing SKUs should confirm the rated compute for the specific model.

For a plant engineer or warehouse automation lead already running a stereo depth pilot, the practical tests are the ones the announcement sets up but does not settle: whether Dynamic Calibration holds depth accuracy across a full shift's temperature swing, and whether Snaps keeps a model honest after the lighting changes in month three. The free Hub tier means the software side of that pilot costs nothing before the first $749 camera arrives.