Skip to content
News Figma Blog Sep 2026

Coinbase used Code Connect to cut AI token costs by 11.5%

Coinbase published results from an experiment in which their design systems team used Claude Code to automate Code Connect mappings across their full Coinbase Design System (CDS) component library. The work took approximately four hours — covering the complete library rather than a curated subset.

The outcomes were measured against a pre-automation baseline: 11.5% reduction in AI token costs, 22.3% faster implementation time for engineers consuming CDS components, and a 22.5% decrease in per-project cost when using AI coding tools alongside the mapped system.

The most concrete improvement the team called out was icon handling. Before Code Connect mapping, AI coding assistants would sometimes hallucinate icon names — generating code that referenced icons that did not exist in the library. After the mapping was in place, the team described this problem as having “disappeared overnight.” The system now gives the AI assistant accurate information about what actually exists in the component library, which eliminates the guessing behavior.

The broader argument in the article is that Code Connect changes what the AI coding assistant can see. Without it, an assistant working in a codebase has no reliable way to know what the design system contains. With accurate mappings, the assistant can reference real component names, real prop values, and real token identifiers, which makes its output more likely to be correct on the first pass.

For design systems teams considering this kind of automation, the Coinbase case is notable because they ran it across an entire production library rather than piloting on a small set. The four-hour automation timeline — using Claude Code to generate the mappings — is the key operational detail for anyone estimating the effort involved.