OpenAI has created a conversational Codex pet powered by GPT-6 Astra.

A researcher didn't want to turn on his laptop, so he asked AI to build a piece of hardware.

The drama of this incident does not lie in "AI building hardware" itself, but in the moving boundary it exposes: AI can now orchestrate a complete workflow from sketches to circuit boards, but it still stops short at tolerance and assembly details. This "Codex Pet" unveiled by OpenAI on November 3, 2026, is exactly the most honest scale mark of this boundary.

One Week, From Sketch to Prototype

The trouble of OpenAI researcher Justin Lin was very specific: he didn't want to open his laptop and whisper to Codex. This complaint gave birth to a palm-sized desktop device — a small screen and a push-to-talk button. He named this prototype "Codex Pet".

What is really noteworthy is not the device itself, but how it was built. The shell, internal structure, circuit board layout, firmware and software integration were all completed by GPT-6 Astra. Two team members used Astra to finish the design in less than a week, and a contract manufacturer delivered a small batch of prototypes in less than a month. According to OpenAI, "there was no direct manual modification" in the design process.

Justin Lin had no prior experience assembling circuit boards from scratch. He described the process as "extremely arduous", and his operation method was summed up in one sentence: "I told Chat what I wanted, gave it some sketches, and then it designed the circuit board."

But OpenAI also admitted the other side: Astra "did not always take into account the production capacity constraints, tolerance requirements and details required for assembly of manufacturers". To make the design production-ready, it still requires "several rounds of review and iteration", as well as "several rounds of communication" with the factory.

This means that what Astra delivers is a logically self-consistent design, not a physically manufacturable one. The gap between the two is exactly the part of traditional hardware engineering that relies most heavily on experience and tacit knowledge. It is true that the design was completed in one week, but between "design" and "can be manufactured", there are still human beings standing.

The Truth of Routing: Orchestration, Not Mastery

In September 2026, circuit board manufacturer NextPCB conducted an independent test, providing a more precise coordinate for understanding Astra's capability boundary.

Hubert Hu, the technical director, asked Astra to complete the routing on a circuit board with completed component layout but no routing. Astra delivered the result in about 12 minutes. Judging only from the start and end points, the conclusion seems very clear: AI has completed PCB routing.

But the execution record tells a different story. Astra's actual operation was: export the design file from KiCad, call the open-source auto-router Freerouting to complete the core calculation, and then import the result back to KiCad. Freerouting is not a new tool, and it has a mature KiCad workflow in the PCB design community. What Astra did was to identify that "this task can be solved with Freerouting", and then accurately execute the multi-step orchestration of export, call and import.

NextPCB's conclusion is therefore much more restrained than the demo video: "Freerouting completed the core routing calculation, while Astra coordinated the entire workflow." Hubert Hu commented that this orchestration capability "is still useful" and may eliminate a large number of manual operations, but it is "technically different from proving that a general large language model has independently mastered PCB routing through visual understanding, mouse operation and keyboard input".

A more rigorous control comes from another set of tests: when Astra was restricted to operate only through the KiCad graphical interface and could not use external tools, it only drew three short traces within an hour, and then Hubert Hu terminated the operation.

This set of comparisons reveals a fact that is easily obscured by the rhythm of the demo: Astra's "physical engineering capability" currently highly depends on whether it can find suitable tools in the existing software ecosystem and call them correctly. This is a real and valuable capability — an agent that can understand engineering tasks, select tools, and coordinate multi-step workflows can indeed significantly reduce manual operations in a real engineering environment. But the distance between it and "AI independently mastering engineering judgment" should not be smoothed out by the smooth demo.

37%: Judgment Remains a Narrow Gate

If the test of NextPCB defines the boundary of "operational capability", then another set of data defines the boundary of "judgment capability".

The engineering evaluation benchmark RWE-Bench was released by Xplore AI on October 27, 2026. The results show that among 20 questions set by practicing engineers that require actual completion of design with CAD and simulation software, GPT-6 Astra's pass rate is 37%, the highest among all participating models, but more than 60% of the tasks still failed the item-by-item verification of the physical solver. The questions cover crankshaft design, DDR3 trace routing, and turbine blade heat transfer. The scoring does not rely on human impression, but on the program checking hard indicators such as stress upper limit, timing deviation, impedance, and heat transfer coefficient.

What is more noteworthy is the answer time: each question allows 24 hours, but the model submits the paper on average in two or three hours, and the longest time taken is only about 4.6 hours. The model tends to "think it has finished". However, the completion of engineering design is never defined by the submission action, but by the satisfaction of physical constraints.

This echoes OpenAI's own disclosed BenchCAD 2.0 data: an updated internal model has an average score of 0.875, while GPT-6 Astra scores 0.594 under the highest reasoning setting. OpenAI did not disclose the details of the internal model, but this set of figures itself shows that the capabilities demonstrated by Astra are far from reaching the upper limit of the current technological frontier. Codex Pet is a credible capability demonstration, but it shows the position of a moving boundary at a certain moment, not the end point.

Why Is It Called "Pet"

Going back to Justin Lin's specific pain point, you will find that the choice of this prototype has a simple rationality.

OpenAI announced at its DevDay in September 2026 that Codex CLI now supports voice dialogue, allowing developers to directly launch and guide programming tasks through voice. Voice is evolving from an additional function to a viable channel for agent interaction. But Justin Lin's complaint is not about voice itself, but about the scenario where voice occurs: turning on the laptop and speaking to the screen is physically still the posture of "working in front of the computer". A small desktop device with push-to-talk does not change the interaction modality first, but the physical position where the interaction occurs — it turns "speaking to Codex" from an action that requires entering a working state into an action that can be completed casually.

This echoes OpenAI's broader hardware narrative. Company president Greg Brockman confirmed in July 2026 that what OpenAI is building is a "family of devices" rather than a single product, and the underlying judgment is that "speaking is more natural than typing". The design cooperation with Jony Ive and LoveFrom, the multi-billion dollar acquisition of io, and the first device scheduled to debut in the second half of 2026, together form a trajectory extending from a software company to a physical product company.

The position of Codex Pet in this trajectory is more like an internal capability reconnaissance: before a product line with huge investment is fully launched, use the minimum cost and the shortest cycle to test how far the model can go on the chain of "from digital design to physical manufacturing" and where it will get stuck. The places where it gets stuck — manufacturing tolerances, assembly details, physical intuition of engineering constraints — are exactly the directions that subsequent model iterations need to target.

The One Who Enters Through the Side Door

Reading this "Codex Pet" as a declaration that "AI has begun to build hardware" may miss its more informative side.

What it truly demonstrates is a layered entry method: AI first establishes capabilities at the operation layer and orchestration layer of digital tools, and then gradually penetrates into the deeper layers that require physical judgment and manufacturing knowledge. At the operation layer, it can efficiently call professional tools like Freerouting to string together multi-step workflows; at the orchestration layer, it can combine the shell, circuit board, firmware and interface into a logically complete design in less than a week; but at the judgment layer — knowing why a trace should be routed this way, why a tolerance should be left this much, and why an assembly sequence cannot be reversed — it still requires repeated correction by human engineers and factories.

The test of NextPCB makes this distinction clear enough: Astra "orchestrated" a routing workflow, but that does not mean it "became" a router.

The cruelty of hardware engineering is that the physical world does not accept "logically correct" designs. An assembly that fits perfectly in the CAD file may not be able to be installed due to the accumulation of tolerances of a few tenths of a millimeter; a circuit board that is connected from the perspective of the routing algorithm may fail completely in the EMC test due to unreasonable return paths. The shortcoming exposed by Astra on Codex Pet — not considering manufacturing constraints and assembly details — is exactly the typical entry point for this kind of "physical world punishment".

And what it has already achieved, those design cycles compressed to one week, automatically connected toolchains, generated firmware and software integration, point to another possibility: the parts of engineering work that can be formalized, orchestrated and accelerated are being steadily stripped out and handed over to the model for processing. The remaining parts — defining what is worth doing, judging what is physically feasible, and taking responsibility for the final consequences — remain firmly on the human side.

Perhaps the most honest revelation of this "pet" is: the way AI enters physical engineering is not to knock on the front door and announce replacement, but to come in through the side door, quietly take over those tedious, repetitive, tool-calling tasks first, and leave engineering judgment untouched until the last moment. It is true that the prototype was built in one week, and the 37% pass rate is also true. The distance between the two numbers is exactly the pit that this industry needs to fill next.

This article is from the WeChat official account "Tech Business", written by Tech Business, and authorized for release by 36Kr.