Can AI design circuit boards yet?

OpenAI recently showcased a demonstration of GPT-6 Astra working on a circuit board within KiCad. While it appears impressive, the technology is far from being able to design an entire phone in a single prompt. This demonstration has sparked an important question: how can we determine if the electronic designs produced by AI are of high quality?

Current AI models have an extensive knowledge of electronics, thanks to their exposure to textbooks, datasheets, application notes, and code. However, when it comes to utilizing this knowledge in conventional design tools, they often struggle. The AI spends a significant amount of time navigating graphical CAD interfaces, keeping track of on-screen elements, and managing application state.

To overcome these limitations, the authors of a benchmark called EEBench have developed a method that allows AI agents to work directly with circuit designs in declarative code. This approach enables the AI to manipulate components, connections, and electrical constraints more effectively. By simulating the circuit, running tests, and inspecting the results, the benchmark focuses on evaluating the actual electronics performance rather than just the graphical operations.

A practical example provided is a residential energy meter circuit. When the 5V supply fails, the circuit must maintain the processor's functionality for 20ms to save accumulated data. Most AI models typically suggest adding a capacitor to solve the issue. However, in reality, selecting the right type of capacitor becomes more complex due to factors like voltage rating, cost, package size, and tolerance variations.

EEBench simulates the power outage and assesses various aspects such as voltage stability, effective capacitance, recovery time, and adherence to component limits.

EEBench takes into account real manufacturer parts and their specifications from datasheets, incorporating them into SPICE models. The AI agent must find a combination of parts that meet the desired electrical performance, are readily available, can be ordered, and fit within a reasonable price range. This task reflects the actual trade-offs faced by electrical engineers, focusing on balancing performance, cost, and supply chain considerations.

The benchmark's measurements are fully deterministic, involving building the submitted design, creating the circuit graph and bill of materials, and executing SPICE simulations and design checks. Each requirement is measured against specific limits, providing a technical score. The final evaluation combines this technical score with cost efficiency, considering the reference bill of materials only after the circuit functions properly.

EEBench V1 covers both analog and digital design aspects through simulation. It does not yet encompass the entire product development process, including layout, manufacturing, and assembly. However, the authors are planning to expand the benchmark to address these additional stages in the future. The current focus on requirements, design, and verification enables the evaluation of engineering work objectively.

In their September 1 results, EEBench V1 showcased promising outcomes. Claude Opus 5 achieved a score of 61.6% across 13 tasks, while Grok 4.6 scored 57.1%, closely followed by Claude Fable 5.1 at 56.4%. This progress is particularly encouraging as frontier labs like xAI are actively incorporating EEBench into their model cards to demonstrate the engineering capabilities of their models.

The rise of Grok 4.6, in particular, is noteworthy. xAI attributes their success to high-quality engineering data and domain-specific reinforcement learning training, including computer-aided design. Such results suggest that EEBench is gaining recognition as a valuable tool for assessing the engineering abilities of AI models.

Comparatively, GPT-5.5 and GPT-5.6 Sol scored 42.3% and 39.4%, respectively, indicating that the performance gap between frontier labs and OpenAI models is narrowing. However, we do not yet have a result for GPT-6 Astra. The authors are eagerly anticipating the performance of GPT-6 Astra on circuit-design tasks after witnessing its ability to work on a KiCad board.

The benchmark's deterministic nature and focus on actual electronic design challenges make it a valuable metric for evaluating AI's progress in the field of electronics engineering.