Event
Vals AI’s SWE-bench Multimodal leaderboard, updated October 13, 2026, ranks Claude Opus 5.5 first with 70.21% accuracy. Claude Fable 5.1 follows at 61.25%, and Claude Sonnet 5.5 at 58.33%.
The benchmark covers 480 tasks from 11 repositories. Each task asks an agent to address a JavaScript software issue whose report depends on a screenshot or short recording, then grades the final code diff with repository tests. Vals AI says its published runs disable internet access because agents with general access can locate the upstream pull request containing the fix. Its page reports that closing network access reduced scores by 21–35 percentage points for three models compared in its diagnostic.