Meta's Muse Code just proved something huge about AI coding agents.
Cline extracted Muse's system prompt and added it to their own harness.
Same model, same task, only the prompting changed:
• 2.7x fewer tokens
• 2x faster
• 2.4x cheaper
Better harnesses > bigger models.
🚨 Keep an eye on DeepSeek V4 Pro! 👀
DeepSeek just dropped V4 Flash 0731 and confirmed the official V4 Pro is right around the corner.
Here is what we know:
🔹 1.6T total / 49B active parameters
🔹 Huge 1M token context window
🔹 V4 Flash is already beating V4 Pro Preview in several benchmarks! 🤯
This is shaping up to be their most powerful AI yet. Factoring in the insane price-to-performance ratio (<$1 per 1M I/O tokens 💸), it could easily go toe-to-toe with GPT and Claude. 🥊
Are you planning to make the switch? 👇