Cody Wilson Claims First AI-Designed 3D-Printed Gun After Three-Shot Test
Cody Wilson says he has built the first AI-made gun, a 3D-printed .22 caliber handgun called the WASP. The pistol was fired in front of a reporter outside Defense Distributed headquarters in Austin, Texas, according to a report published Oct. 16, 2026 The Verge.
The WASP misfired on the first trigger pull. It fired on the second attempt and hit a cardboard target. Wilson fired it handheld on a third attempt, and that shot missed high.
The claim comes 13 years after Wilson built the first 3D-printed firearm, per the same report.
The design tool, in Wilson's account, is a custom model called GatGPT. It is described as an AI that can generate guns. It was trained on the DEFCAD library of gun files and answers questions about firearms.
DEFCAD was launched in 2012 and is led by Wilson. It is described as the world's largest online repository of digital gun and component files.
The broader context here is a shift from general-purpose assistants to smaller, specialized tools trained for one job. That pattern is familiar in software, where a model trained on a focused set of good examples can draft usable work faster than a general model starting cold. With physical objects the draft still has to survive physics, small differences in size and printing variation. A generated shape is only a suggestion until it is printed, assembled and tested under controlled conditions. Here the training files and the distribution site are run by the same group, which shortens the loop between finding a file and generating a new one.
In my view, the three-shot sequence in Austin is more telling than the first AI-made label. Misfire, hit, miss is a normal early hardware cycle, and it shows generation has not removed trial and error. There is a gap between parametric CAD, which is design software that records intent, limits and failure modes, and token prediction, which predicts likely shapes from patterns in data. Closing that gap takes test measurements fed back into the model, not just more files added. Reliability comes from measured pressure, tight size control and repeated testing, not fluent output.
Looking at what this means in practice, the more interesting piece may be GatGPT rather than the WASP itself. A question-and-answer tool over a large library of printable parts changes how people find components. A generative layer on top changes how they draft new ones. Neither removes the need to verify. Worth flagging, a .22 caliber handgun is a low-energy place to start that learning process. It keeps variables down while the toolchain matures, which makes sense when the design source works on probabilities.
The long arc here points toward cheaper iteration for physical goods. Inexpensive printers, shared libraries and now specialized models cut the time from idea to testable object. My kids grew up assuming software could be reshaped at will. Their assumption about hardware is starting to look similar. That expectation will produce many failed prints before it produces dependable systems, but the direction is consistent. The work ahead is measurement.
