DeepSeek V4 Pro: 1.6T Open MoE Model Hits 80.6% SWE-bench at One-Tenth GPT-5.5's Price

Summary: DeepSeek launched V4 Pro on April 24, 2026 — repeating the V3 playbook at trillion-parameter scale by pairing near-frontier coding performance with open weights and aggressive pricing.

Key Facts

• Architecture: 1.6T total / 49B active parameters (Mixture-of-Experts); 1M-token context window; up to 384K output tokens; thinking and non-thinking modes
• Pricing: $0.435/M input, $0.87/M output — approximately one-tenth of GPT-5.5's output price
• Benchmarks: 80.6% on SWE-bench Verified, placing it among the top coding models available
• License: MIT — fully open for commercial use and self-hosting; still in preview as of June 2026

Why It Matters

DeepSeek V4 Pro proves the V3 thesis at a larger scale: open weights, frontier-class coding capability, and commodity pricing — available for on-premises deployment. That combination puts direct cost pressure on OpenAI and Anthropic's enterprise offerings and lowers the barrier for teams that want powerful coding AI without vendor lock-in or per-token API costs.

Further Reading

• DeepSeek V4 Preview Release Notes — DeepSeek Official
• DeepSeek V4 Architecture Deep Dive — MorphLLM
• DeepSeek V4 Pro Complete Guide — AI Made Tools