GPT-6 Astra for PCB design
GPT-6 Astra can operate a PCB tool on your desktop, which is genuinely new. It still cannot route a board, because routing was never the part a language model was doing.
Short answer
GPT-6 Astra can drive KiCad or another EDA tool through your desktop and click through a design the way a person would. It does not contain a routing engine or a design rule checker. When it presses Auto-Route, the tool's router does the work — the same millions of search-node expansions, with the same result, whether a model asked for it or you did.
See how PCBEditor does it →What GPT-6 Astra does well
- GPT-6 Astra can operate desktop EDA software directly — open KiCad, place parts, press Auto-Route, export files. That is a real capability the earlier chat models did not have.
- It is strong at the judgement half of hardware design: choosing a regulator, reading a datasheet, explaining why a decoupling capacitor belongs close to a supply pin.
- It can chain a multi-step workflow — draft a circuit, drive the tool, inspect the output, and try again — which is where agentic systems genuinely earn their cost.
- At 99.9% on ARC-AGI-3 it is markedly better at recovering from its own mistakes mid-task than the models before it.
What it cannot do, and why
- It has no routing engine of its own. Driving a tool means the TOOL routes; a better model does not make that router faster or more complete.
- It cannot shorten the compute. Routing is a constrained search over physical space — an eight-layer board with a large BGA can take millions of node expansions and still finish incomplete. None of that work is language, so none of it improves with model quality.
- It cannot verify manufacturability by reasoning. A design rule check is a set of measurements against thresholds, computed rather than recalled.
- It cannot make a GUI faster than an API. Clicking through a desktop tool is the slowest possible way to drive an engine that could be called directly.
The split that actually works
The useful way to think about it is not GPT-6 Astra versus a PCB tool. It is which half of the problem you are asking a language model to solve. Choosing parts, reading a datasheet and deciding what the circuit should do are judgement tasks, and a model is good at those. Placing components, routing copper and proving a board passes design rules are geometry and constraint problems, where the answer is either correct or it is not.
PCBEditor is built on that split. A language model chooses parts and interprets what you asked for. A purpose-built router and a rule engine produce and verify the board. Correctness does not depend on which model is behind the pipeline, which is the only way the output can be trusted.
Using GPT-6 Astra to drive a PCB tool? Give it one with a real router behind it: deterministic placement, multi-layer routing, DRC and Gerbers, driven from the same plain-language request.