DeepSeek releases updated V4 Pro model, boosting coding capabilities

DeepSeek released version 0813 of its flagship V4 Pro model, aimed at complex tasks, programming, and AI agent work, as stated in company documentation. On the independent Artificial Analysis platform, the new version scored higher than Z.ai's GLM-5.2, one of China's strongest coding AI models. According to DeepSeek, V4 Pro scores improved notably in internal tests: in Terminal-Bench 2.1 (command-line tasks) from 72.1% to 87.9%, in CyberGym (vulnerability handling) from 52.7% to 83.3%, and in DeepSWE (programming) from 12.8% to 62.7%. The model has 1.6 trillion parameters with 49 billion active, a context window of 1 million tokens, and maximum output length of 384,000 tokens. Pricing is $0.435 per million input tokens (without cache) and $0.87 per million output tokens. V4 Pro was initially introduced in April 2026 in the chatbot and on Hugging Face; the updated version released August 13 is available via the app, web version, and tools like Codex, Claude Code, and OpenCode. Earlier in August, DeepSeek announced plans to raise prices, and Bloomberg reported the company is building a large data center in northern China.