DeepSeek V4 remains on track for a late-July release after slipping from a mid-July target, with limited gray testing underway for Pro and Flash variants focused on reasoning, speed and agent functions.
DeepSeek V4 is still expected to be released by the end of July after missing its earlier mid-July target, with limited gray testing now underway for DeepSeek-V4-Pro and DeepSeek-V4-Flash. The update is focused on better reasoning efficiency, faster responses and stronger agent capabilities. Earlier reporting said DeepSeek-V4 was being benchmarked against Anthropic Opus 4.8 and described as rivaling OpenAI GPT-5.6 Sol in code generation, while also improving performance in 3D modeling, SVG image generation and mini-game development. DeepSeek had also completed a CUDA-free adaptation for V4 to support a wider range of domestic AI chips, and the platform was expected to introduce peak and off-peak API billing with peak-hour rates at double off-peak levels. Separate earlier reporting indicated the Harness team may launch its first product alongside the V4 release and that a coding agent was being benchmarked internally against Claude Code.