Which models can actually do CAD?

Read this gap with care

This run limited Kimi K3.1 Pro to 32,000 output tokens per response. On the articulated figure task, two of its three attempts reached that limit in their first response and built nothing. That one task is most of the gap between the two models: without it, both score about 92.

We then raised Kimi's limit to 60,000 and ran that task six more times on its own. All six passed the gates and none reached the limit. The task's mean score rose from 29 to 89, against 98 for Sonnet 5.5. The attempts also cost more and took longer: $1.30 and about 12.7 minutes each, up from $0.54 and about 5 minutes.

Those six attempts are not in the numbers on this page. Using their score for that task would put Kimi at about 91.6 against Sonnet's 92.4, but that is an estimate, not a measured run.