GPT-6 "Astra" brings vibe coding to hardware: this is how it designs complete PCBs in KiCad

In early 2025 the term "vibe coding", coined by OpenAI co-founder Andrej Karpathy, took off as a way to describe how language models made it possible to program software by describing in plain English what you wanted to build. Until now, improvements in these models had been concentrated in text-based reasoning, image recognition and computer use in a generic sense. OpenAI's release of GPT-6 "Astra" adds a new dimension to that "computer use" capability: directly operating computer-aided design (CAD) tools, and specifically the software used to design printed circuit boards (PCBs), the physical foundation of virtually all modern electronics.

The difference from traditional vibe coding is significant: in software, the model generates text that is then turned into executable code. With Astra operating a PCB design program, the model is no longer limited to producing a description or a file: it interacts directly with the same tool a human engineer would use, placing components and routing the connecting traces inside the CAD environment itself. This new practice is starting to be called "Vibe CAD Engineering": describing a board, its components, its constraints and its purpose, and letting an AI agent do the design work inside the relevant software.

After the launch, the community on the social network X was quick to put these capabilities to the test, with results that, according to the users who shared them, surpassed what OpenAI itself had shown in its official demo. One of the most widely circulated examples came from a user who, in roughly half an hour, got a version of the model (identified as "Astra-Medium") to design a Bluetooth wearable with built-in artificial intelligence: it included two microphones for "beamforming" (shaping the sound beam), an inertial measurement unit (IMU), an LED and a button. To do so, the model relied on a set of agentic skills called KiStack, designed specifically so that a language model can operate the open-source PCB design software KiCad, together with a derived tool called T3CAD.

Another documented case was that of a user who asked Astra to design, end to end, the PCB for an FPV (first-person view) drone. Starting from an existing reference board, the model completely redid the component placement and trace routing, producing a design described as considerably cleaner than the original. What is striking about this case is that it did not stop at simulation: the design was sent for actual manufacturing (using the EasyEDA tool), surface-mount (SMT) assembly was completed, and the user himself shared photographs of the finished physical board, pending final power-on tests and integration with the PX4 flight control software.

A third example cited is a six-layer board, designed and routed entirely by Astra, which was also sent to a manufacturing contractor and passed pre-production review. The person who shared this case on X enthusiastically framed it as a sign that the era of artificial general intelligence (AGI) is drawing closer, a comment that reflects the community's enthusiasm more than any verified fact, but which illustrates the impact these demonstrations are having.

OpenAI itself, on the GPT-6 "Astra" product page, shows an example of its own in which the model carries out a board design inside KiCad: it takes the component list, reads the layout specifications, places each part in position, optimizes the overall layout and routes the necessary connections to produce a working design, much as a human engineer would. After that process, an engineer is still needed to run the relevant checks to confirm the design works correctly before moving to manufacturing.

To understand why this is a leap beyond traditional CAD tools, it helps to explain how they work today. Standard PCB design software already includes auto-placement and auto-routing functions, but they rely on predefined mathematical algorithms that treat the board as a canvas with several interconnection layers. Because those algorithms lack any real understanding of the three-dimensional physical world, they frequently produce suboptimal results that are poorly suited to manufacturing. This is where the underlying difference comes in: a model like Astra combines its ability to reason about components and their placement with the dexterity to operate the software tools directly, achieving a workflow far closer to a person's.

A central part of that workflow is the Design Rule Check (DRC), an engine present in practically all modern PCB CAD software that examines the layout, the component placement and the board's electrical and physical properties; for example, whether a component makes electrical sense in its position, or whether different layers might overlap and cause problems. Traditionally, these checks flag errors at the end of the process, which an engineer then fixes manually before sending the design to manufacturing. But the DRC is only one part of validating a PCB: a design can pass its basic rules and still require additional validation of signal integrity, thermal behaviour, power delivery, component availability, electromagnetic compatibility and performance under real-world conditions. This is where the engineer's role changes: rather than disappearing, it shifts towards defining what the product must achieve and checking whether the final design genuinely meets those requirements, while AI takes on more and more of the execution inside the CAD environment itself.

For an AI agent, this framework of explicit rules is especially favourable terrain: in theory, it could place and route components, run the DRC, read the results and correct whatever fails, repeating that cycle almost indefinitely until it arrives at a perfect design. In practice, that loop is limited by available compute and stops once it reaches the goal set out in the user's original instruction, rather than running indefinitely. OpenAI, it is explained, has optimized Astra through tool calls and internal reasoning precisely so that it performs well on tasks with many rules to follow, as is the case with PCB design.

That optimization rests on dedicated reinforcement learning (RL) environments, in which the model's agentic use is rewarded when it completes tasks correctly and penalized when it goes off track. This type of training has become so important that labs now devote a considerable amount of compute to this phase alone, which comes after pre-training and fine-tuning, with the aim of steering the model's intelligence towards applying its capabilities correctly across a wide range of tasks. Cited as a public reference point for this trend is the case of Xiaomi's MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL models, whose training showed that scaling real-time RL environments demands a great deal of compute but produces significant performance gains on the usual benchmarks. If a company like Xiaomi makes this public, it is reasonable to assume that labs such as OpenAI have carried out considerably larger-scale RL optimization on their own models, although the exact share of compute devoted to pre-training, mid-training and post-training at the big labs is not known and can only be estimated.

Looking ahead, models can be expected to keep expanding their RL environments in order to bring computer use to a virtually flawless level, which could turn tasks such as PCB design into something almost as trivial as vibe coding an app is today. One prerequisite will remain, however: describing a board well in plain English requires a certain knowledge of hardware engineering and of how an electronic product is specified. Creating a simple PCB for a Bluetooth accessory or an FPV drone is only part of the story; the real test will be whether these systems manage to scale from relatively simple boards to far more complex products, such as motherboards, GPU cards, networking hardware or accelerator boards for AI servers, which in the real world have very high layer counts, tens of thousands of components and demanding power delivery requirements. Only if AI models become viable in those production-grade scenarios could broader commercial adoption of this technology be expected; for now, their real reach is to allow anyone who already vibe codes software to become a "vibe hardware engineer" as well, provided they know how to describe precisely, in natural language, what they want to build.