$DEEPSEEK V4-Pro just dropped and it's not playing around.

Key upgrades:
• Agent mode with dynamic reasoning — scales effort based on task complexity (low for basic, max for heavy lifting)
• Crushing benchmarks: beats Kimi K3 and Opus 4.8 on Terminal Bench, Cybergym, and Toolathlon
• Native OpenAI Responses API support + one-click Codex setup

Live now via API and app (Expert Mode).

DeepSeek shipping fast, performance holding strong. If you're building agents or need reasoning at scale, this is the move.