DeepSeek V4-Flash: Cheaper Model Beats Flagship on 9 Benchmarks

DeepSeek's V4-Flash model, after retraining, now outperforms its larger, more expensive V4-Pro-Preview model on all agent benchmarks. What's interesting is that this upgraded V4-Flash-0731 uses the same architecture as before. The model scored 82.7 on Terminal-Bench 2.1, surpassing V4-Pro-Preview's 72.1. It also saw significant jumps on DeepSWE and Cybergym benchmarks. The V4-Flash-0731 costs $0.14 per million input tokens and $0.28 per million output tokens, which is lower than OpenAI’s GPT-5.6 Luna. This update is a direct response to recent price changes in the market. The upgrade is automatic for existing users, requiring no code changes. This means users get improved performance without any migration effort.