Can AI design circuit boards yet?

We sewage beautiful excited yesterday erstwhile OpenAI put a demo of GPT-6 Astra moving connected a circuit committee successful KiCad connected the beforehand page of its motorboat post. It is cool to spot electronics show up successful a awesome exemplary merchandise for illustration this.

We are evidently still immoderate region from asking an AI to build an full telephone successful 1 prompt. The demo does raise a mobility we person been reasoning astir for a while, though: really do we measurement whether the electronics an AI produces are really immoderate good?

The models cognize a astonishing magnitude astir electronics

Our acquisition has been that existent models cognize overmuch much astir electronics than their output in accepted creation devices tends to show. They person publication textbooks, datasheets, exertion notes and a batch of code.

You tin person an supplier run a graphical CAD tool, but it spends a batch of clip clicking astir and keeping way of what is connected screen. A batch of its discourse consists of coordinates, menus and exertion state.

EEBench uses atopile instead. The circuit lives successful declarative code, truthful the supplier tin activity straight connected components, connections and electrical constraints. It tin alteration the design, build it, tally a simulation and inspect what grounded without leaving the project.

This has worked overmuch amended for america than asking a exemplary to tie lines successful a GUI. It besides intends the benchmark tin walk little clip testing machine usage and much clip testing electronics.

The existent world is messy

One of the nationalist tasks is based connected a residential power meter. When its 5 V proviso disappears, the circuit has to support the processor live for different 20 sclerosis truthful it tin prevention the accumulated reading. The protected rail must enactment supra the processor's 3.0 V brownout period during that window.

Most models intuitively jump to the correct guidelines conclusion: adhd a capacitor.

A existent capacitor makes the task much interesting. A ceramic portion whitethorn supply overmuch little than its advertised capacitance erstwhile it has voltage crossed it. Parts person tolerances. Adding much capacitance costs more, takes up abstraction and makes the obstruction slower to recharge erstwhile the powerfulness returns. A creation that useful pinch nominal values tin neglect pinch the parts that arrive.

EEBench cuts the input powerfulness successful simulation and measures what happens. It checks the voltage passim the outage, the effective capacitance astatine the operating point, the betterment aft powerfulness returns and the limits on package, dielectric, voltage standing and cost.

The metre is 1 of the easier tasks. In a harder analog task, the supplier whitethorn person to synthesize a multiple-feedback low-pass select astir an op-amp, lick the resistor and capacitor ratios for the required poles, and support its gain, cutoff wave and Q wrong their limits aft each constituent is pushed to a worst-case tolerance corner. The harness rebuilds the SPICE platform for those corners, runs the AC and transient captures, binds measurements to named probes, and records each consequence against its little and upper specification limits.

But getting the equations correct is only portion of electronics engineering. EEBench uses existent manufacturer parts, pinch specifications extracted from their datasheets and carried into the SPICE model. The supplier has to find a operation that useful crossed those tolerance corners while besides choosing parts that exist, tin be ordered and are reasonably priced for the product. That trade-off betwixt electrical performance, costs and proviso is overmuch person to designing existent hardware than picking perfect values from a textbook.

This is the portion we find astir interesting, because it is what electrical engineering yet boils down to, conscionable for illustration each different engineering discipline: trade-offs.

How the grading works

EEBench checks are afloat deterministic. It builds the submitted design, constructs the circuit chart and measure of materials, and runs a group of SPICE simulations and creation checks. Each request produces a measurement pinch a limit.

For the energy-meter task, the harness measures the protected obstruction while the input drops retired and returns. Other tasks measurement gain, thresholds, ripple, transient consequence and behaviour astatine component-tolerance corners. The method people is mixed pinch costs ratio against a reference measure of materials. Cost only helps erstwhile the circuit works.

This is akin to giving a coding supplier a compiler and tests, isolated from the tests are measuring voltages and constituent behavior..

EEBench V1 covers analog and integer creation done simulation. It does not yet show america whether a exemplary can laic out, manufacture and bring up a complete product. We want to adhd those parts later. The existent benchmark concentrates connected the requirements, creation and verification loop because that is wherever we tin already grade useful engineering activity objectively. The afloat methodology and sample result explorer are public.

What we are seeing connected the leaderboard

The September 1 results are encouraging. Claude Opus 5 scored 61.6% crossed the 13 tasks successful EEBench V1. Grok 4.6 came 2nd astatine 57.1%, conscionable up of Claude Fable 5.1 astatine 56.4%. A fewer months agone we would not person expected models to do this well.

There was different consequence we were particularly happy to see: xAI included EEBench successful the Grok 4.6 model card. It appears successful the conception connected “engineering acceleration,” alongside evaluations for 3D modeling and parametric CAD. Their published tally put Grok 4.6 astatine 60.0% pinch xhigh reasoning effort. Seeing a frontier lab usage EEBench to picture a caller model's engineering expertise makes america deliberation this is becoming a class people attraction about.

Anthropic's models person consistently done good successful this environment. Grok's emergence is besides interesting. In its Grok 4.6 motorboat post, xAI says the exemplary received high-quality engineering information and RL training successful domain-specific environments including computer-aided design. Its EEBench consequence fits that story.

The OpenAI models we person tested truthful acold beryllium further down the table. GPT-5.5 scored 42.3%, while GPT-5.6 Sol scored 39.4%. We do not person a GPT-6 Astra consequence yet. After seeing it activity connected a committee successful KiCad, we would really for illustration to find retired really it handles these circuit-design tasks.

Training environment

Once we had a simulation harness that could people a circuit, we besides had the beginnings of an RL environment for electronics. The aforesaid checks tin beryllium utilized arsenic reward signals during post-training.

A grounded tally contains useful information. We tin spot which voltage missed its limit, which operating corner failed, aliases whether the exemplary solved the problem pinch an unnecessarily costly design. That gives a training loop much to activity pinch than a exemplary saying that a schematic looks plausible.

EEBench is the small, nationalist position of this work. We are besides starting to activity straight pinch frontier labs that want to make their models amended astatine electronics. If you activity connected evals aliases post-training, you tin talk to america astir the larger information suites and simulation-backed training environments we provide.

So, tin it creation a circuit board?

For a useful and increasing group of circuit problems, we deliberation the reply is already yes. The scores besides make it clear that location is plentifulness near to do. OpenAI choosing a PCB for 1 of Astra's first demos and xAI publishing EEBench successful a exemplary paper some consciousness for illustration early signs of the aforesaid thing: AI labs are starting to return electronics seriously.

There whitethorn beryllium different information constituent very soon. Elon Musk has said that Grok 4.7 is coming wrong weeks aft further training connected a ample postulation of SpaceX data, pinch the purpose of making it particularly bully astatine engineering. The exemplary is not retired and the schedule whitethorn move, but if it arrives arsenic described, we will beryllium keen to put it connected EEBench.

It looks for illustration we are successful for an absorbing fewer weeks. We will support adding harder tasks arsenic the models improve, and we are looking guardant to seeing really Astra and the adjacent procreation perform.

You tin try the aforesaid attack successful atopile today. Give the supplier a committee you person been meaning to build and spot really acold it gets.

So, tin AI creation circuit boards yet? Some of them, yes. We still would not inquire it to creation a pacemaker and blindly instal the result. But we are connected the measurement there.

EEBench is built and funded by the squad down atopile. We salary for the public benchmark runs and do not waste benchmark scores.