EEBench, a new benchmark for evaluating AI on circuit board design, shows current models can handle real electronics tasks but with significant gaps. Claude Opus 5 leads the leaderboard at 61.6% across 13 design problems, followed by Grok 4.6 at 57.1% and Claude Fable 5.1 at 56.4%, while OpenAI's GPT-5.5 and GPT-5.6 Sol score 42.3% and 39.4% respectively. xAI's published run of Grok 4.6 with xhigh reasoning effort achieved 60.0%.
The benchmark sidesteps the GUI-clicking problem that has plagued earlier AI-assisted EDA work. Rather than asking models to operate KiCad or similar tools through screen coordinates and menus, EEBench uses atopile, a declarative circuit description language. Agents work directly on components, connections and electrical constraints, then build, simulate and inspect failures without context switching. This approach lets the benchmark measure electronics knowledge rather than computer-use ability. According to the EEBench team, "current models know much more about electronics than their output in conventional design tools tends to show," having trained on textbooks, datasheets and application notes.
The tasks themselves reflect real engineering trade-offs. One public problem, based on a residential energy meter, requires a circuit to keep a processor alive for 20 milliseconds after power loss using a capacitor hold-up rail. The benchmark does not accept nominal-value designs; it simulates ceramic capacitor behavior under voltage, component tolerances, cost and supply constraints. A harder analog task asks agents to synthesize a multiple-feedback low-pass filter, solve resistor and capacitor ratios for required poles, and keep gain, cutoff frequency and Q within limits across worst-case tolerance corners. The grading harness runs SPICE simulations at those corners, measures voltages and component behavior against specification limits, and combines technical score with cost efficiency against a reference bill of materials.
The leaderboard results show both capability and distance remaining. Claude Opus 5's 61.6% means it fails on roughly four in ten tasks. EEBench reports that xAI included the benchmark in the Grok 4.6 model card under "engineering acceleration," suggesting frontier labs are beginning to treat circuit design as a measurable capability. According to xAI's Grok 4.6 launch post, the model received "high-quality engineering data and RL training in domain-specific environments including computer-aided design," which appears reflected in its performance.
EEBench V1 covers analog and digital design through simulation but does not yet measure layout, manufacturing or bring-up. The benchmark focuses on the requirements, design and verification loop because those are the stages where objective grading is already possible. The EEBench team states that it is "starting to work directly with frontier labs that want to make their models better at electronics," offering larger evaluation suites and simulation-backed training environments beyond the public benchmark.
The practical limit is clear: the source says "we still would not ask it to design a pacemaker and blindly install the result." For teams considering AI-assisted EDA workflows, EEBench provides the first quantified measure of whether a model can actually solve circuit problems or merely talk about them plausibly. If your design domain matches the benchmark's scope—analog and digital circuits where simulation can verify correctness—the leaderboard scores tell you what to expect. If it does not, you have no measurement yet.