Mistral Vibe, Claude Code, Cursor, Codex — agent benchmark on PR task
Benchmark compares four code agents on a single scaffold-to-pull-request workflow.
• Mistral Vibe: open weights, self-hostable, cost-efficient async
• Claude Code: proprietary, highest context, production-grade reliability
• Cursor: IDE-native, real-time, local reasoning
• Codex: legacy baseline, GPT-3.5 class, reference only
Trade-offs: openness and cost versus depth of reasoning and IDE integration.