NVIDIA and chip-design software developer ChipAgents launched Build-a-Chip-Bench; the best-performing agent, powered by Anthropic Fable 5.1, scored 16.6%.

The benchmark evaluates autonomous chip design and implementation. Supplied launch messages identify the joint NVIDIA–ChipAgents benchmark as Build-a-Chip-Bench. A separate report attributed to The Information says it measures how different AI models perform on chip-design tasks.

The strongest reported result was a 16.6% score. The best-performing agent was powered by Anthropic’s Fable 5.1. The supplied report does not specify the task count, scoring formula, or evaluation budget, so the result should not be reinterpreted as a percentage of commercial chips successfully designed.

ChipAgents CEO William Wang described the benchmark as covering the complete chip-design workflow. He said this made it harder than existing benchmarks. This explanation is Wang’s characterization of the benchmark, rather than a supplied side-by-side validation against other evaluations.

The report also provides limited commercial background on ChipAgents. Wang said its customers include several semiconductor-design companies with multibillion-dollar market capitalizations, but declined to name them. He also did not disclose whether NVIDIA was an investor or customer; the confirmed relationship here is the benchmark collaboration.