AI PCB Design Enters a Measurable Phase
AI PCB design is no longer only a demonstration concept. EEBench has created a public evaluation around analog and digital circuit tasks, asking a more practical question than whether a model can operate KiCad: does the resulting circuit satisfy explicit engineering requirements? OpenAI’s demonstration of GPT-6 Astra working in KiCad has increased attention, but the current evidence is still concentrated on simulation and design verification rather than complete product delivery.
How EEBench Tests Circuits Beyond Their Schematics
EEBench represents a design as declarative code, allowing an agent to work directly with components, connections, and electrical constraints. A submission is built into a circuit graph and bill of materials, followed by deterministic SPICE simulations and design checks. The evaluation can measure voltage, gain, thresholds, ripple, and transient response, comparing each result with specification limits. A schematic that merely looks plausible is therefore not enough for a strong result.
A residential energy-meter task illustrates the difference. When its 5-volt input disappears, the circuit must keep the processor alive for 20 milliseconds while maintaining the rail above the 3.0-volt brownout threshold. Models commonly identify the need for a capacitor, but a real ceramic capacitor may provide only part of its nominal value under bias. Tolerance, voltage rating, package, cost, and recharge behavior also matter. In one submission, a nominal 22-microfarad capacitor delivered only 11.4 microfarads at 4.7 volts in simulation, and the rail fell below the requirement after 0.85 milliseconds. The design built successfully but still failed its purpose.
From Ideal Values to Component Trade-offs
Harder tasks ask agents to synthesize op-amp filters while keeping gain, cutoff frequency, and Q within limits, then rerun AC and transient analyses at worst-case tolerance corners. EEBench also uses real manufacturer parts and carries datasheet specifications into its SPICE models. Success therefore requires more than solving textbook equations. The agent must find a combination that exists, can be ordered, is reasonably priced, and continues to meet its limits across operating conditions. That balance among electrical performance, cost, and supply is closer to practical hardware engineering.
Results Show Progress and Clear Boundaries
In the September 1 leaderboard reported by the source, Claude Opus 5 scored 61.6% across 13 tasks, followed by Grok 4.6 at 57.1% and Claude Fable 5.1 at 56.4%. These results show that models can already complete useful work on a growing set of circuit problems, but they do not mean that a model can design a complete production board. EEBench V1 mainly evaluates analog and digital design in simulation; full layout, manufacturing, and physical bring-up are not yet covered. Grok 4.6’s model card cites EEBench, suggesting that engineering evaluations are becoming part of model reporting, although settings and results still require careful comparison.
For engineering teams, simulation-driven agents appear suited to design exploration, constraint checking, and failure diagnosis, especially when each requirement can be measured explicitly. They should not replace experienced review of safety, reliability, EMI, thermal behavior, manufacturability, or physical prototypes. EEBench’s most valuable direction may be turning failures into learnable feedback: which voltage missed its limit, which tolerance corner failed, and whether the design paid unnecessary cost to pass. The next meaningful test is whether models can move from declarative circuits and simulation results into a loop that includes layout, manufacturing, measurement, and iteration.
Conclusion: Capable on Some Circuits, Not Ready for Blind Delivery
The careful answer is that some models can already design a useful class of circuits, but there remains a clear gap between designing a circuit and delivering reliable hardware. EEBench makes that gap measurable and shows why a successful build is not the same as functional success. As evaluations add more real components, operating corners, and downstream hardware steps, the boundary of AI PCB design will become clearer than any single demonstration can show. For now, the sensible role is to use models inside a verifiable engineering loop while retaining final responsibility for human review and physical validation.