OpenAI just dropped a bomb in the AI industry, and as per the demo they have shown, they illustrated a few people talking to GPT to get their tasks done, like creating a 3D model using CAD software, building a game with that 3D model, and accessing the slicer on your desktop to slice it and send it for 3D printing. Sounds fascinating, right?
Now let's look at the closer view. I have used Antigravity by Google, Claude Code, and Codex with GPT-5.6 and older models. One thing I have noticed is that the more agentic the system is, and the more iterations it does, the more work it gets done for you.
GPT-5.6 Sol costs $4/1M input and $20/1M output. It works for me for things like FE testing, browser control, and some desktop software control as well. So what do you get with GPT-6 Astra at $10/1M input and $50/1M output? That is a flat 2.5x across input, cached input and output. And on ARC-AGI-3 it scores 99.9%, against 30.2% for Claude Opus 5 and 7.8% for GPT-5.6 Sol.
Let's be fair about that price: it's a good deal. Going from 7.8% to 99.9% on a reasoning benchmark for 2.5x the token cost is not expensive, if reasoning is your bottleneck.
The footnote on that 99.9% is the interesting part
ARC Prize published the conditions, and they are worth reading carefully.
The 99.9% comes from OpenAI's Provider Adapter harness, which preserves opaque reasoning state between requests so the model can reuse prior work. Cost of that run: $18,817.
Run the same model through the Standard harness, which is stateless and provider-neutral, and it scores 62.7%. That run cost $26,098. More money, a third of the score.
Same model. Same weights. The difference is entirely in the scaffolding around it.
That is the most useful fact in the whole launch, and it is not the one on the chart. It says the system you wrap a model in can be worth more than 37 points of benchmark, which is the same reason a purpose-built tool beats a general model clicking through someone else's GUI.
And on computer use specifically, the numbers are real: 72.6% on OSWorld 2.0 against 65.7% for Sol, finishing tasks in about 40 minutes where the previous model took 75. Astra genuinely operates desktop software better than anything before it.
But operating a router is not being a router
Point that same model at a PCB and reasoning isn't the bottleneck. Geometry is. You are paying 2.5x for a much better operator of a tool that still has to do the routing itself, and routing is not a reasoning problem.
So can it generate an end-to-end PCB?
The answer might be yes or no, because the question is wrong.
It's a great model that can use your desktop tools, but it doesn't have a natively built PCB generation engine or a CPU-heavy routing engine. It's just good at using tools. Whereas AI-native tools like PCBEditor are built to generate PCBs using AI. So yes, maybe you can use GPT-6 Astra to generate PCBs by driving tools like KiCad and PCBEditor.
What "CPU-heavy" actually means
Routing a board is a constrained search through physical space. A router has to find a path for every net across a grid of the whole board, on every layer, while respecting clearance to every other piece of copper, and then undo and retry when a later net blocks an earlier one. On a dense multi-layer board with fine-pitch parts, that runs to millions of grid evaluations and takes minutes of real CPU time.
None of that work is language. A model that reasons better does not reduce the number of cells a router has to evaluate, because the router is not reasoning. It is measuring.
And here's the thing: a model driving KiCad through a GUI still has to call KiCad's router. That router does the same millions of evaluations whether a model asked for it or you clicked the button yourself. The intelligence and the computation are separate costs, and only one of them is billed per token.
About that PCB Generation in Demo
The image published by OpenAI is a screenshot of a KiCad desktop window. It reads as though the model generated a PCB, when what it shows is a model operating a tool, the same way you, I, or any other desktop-controlling AI model would.
That distinction matters more as these models get better, not less. Astra is genuinely the best desktop operator that has shipped. The engine underneath the tool is a separate problem, and it will not be solved by the next model either, because it was never a model problem.
Where would you put the line? If you have built something where a frontier model drives a domain tool, I would like to know where the model stopped being the hard part. For me it was the moment the search space stopped fitting in a context window and started fitting in a CPU.
Top comments (0)