18 September 2026
Analog is silicon's biggest bottleneck. GenAlpha builds AI Designers that accelerate it. Agents that do the design work in a real flow, drawing schematics & layout, running & debugging DRC, LVS, and parasitics.
Our AI Designers can be paired with any of the leading models, so we evaluate those models regularly — which ones do the best analog work, and how much output and time each one takes to get there. Our primary evaluation framework is GenAlpha AnalogBench: a curated set of real analog design and layout tasks, executed by our AI Designers paired with each model. We recently evaluated twelve leading open- and closed-weight models, and want to share the results.
Higher scores are better.
Highest score in each column in teal. Simple edits, such as deleting an instance or moving a net, are counted with layout.
| Model | Overall | Schematic | Layout | Q&A |
|---|---|---|---|---|
| claude-fable-5.1 | 796 | 794 | 765 | 856 |
| claude-opus-5 | 771 | 798 | 695 | 862 |
| grok-4.6 | 715 | 666 | 737 | 757 |
| gpt-5.6-sol | 709 | 710 | 666 | 787 |
| gemini-3.8-flash | 701 | 749 | 637 | 737 |
| muse-spark-1.3 | 681 | 694 | 657 | 702 |
| gpt-6-astra | 665 | 867 | 430 | 748 |
| glm-5.3-flash | 624 | 614 | 585 | 717 |
| kimi-k3 | 621 | 524 | 603 | 819 |
| qwen3.8-max | 616 | 622 | 532 | 753 |
| deepseek-v4.1-flash | 528 | 374 | 572 | 717 |
| inkling | 184 | 125 | 160 | 329 |
One point per model. Higher and further left means a better score for less work. Tokens and time include the AI Designer and any sub-agents it starts.
Generation Alpha Transistor · San Francisco, CA