The production build of DeepSeek’s V4 Pro, tagged 0813, hit general availability on August 12, ending a preview that ran nearly four months.
The model is now live on OpenRouter and DeepSeek’s own API with unchanged pricing: $0.435 per million input tokens on a cache miss, under a cent per million on a cache hit, and $0.87 per million output tokens. The model keeps a one-million-token context window and allows up to 384,000 output tokens.
Behind the price is an efficiency story. Its mixture-of-experts design runs 1.6 trillion total parameters with 49 billion active per token, and DeepSeek’s two attention variants cut per-token inference compute to 27 percent and KV cache to 10 percent of the V3.2 generation at the million-token setting. Pretraining covered more than 32 trillion tokens.
Self-reported scores at maximum reasoning effort include 80.6 percent on SWE-bench Verified and 67.9 percent on Terminal Bench 2.0, with Humanity’s Last Exam at 37.7 percent. The open-weights Pro repository passed 1.4 million downloads on Hugging Face over the last month.
The series began with an April 24 preview under the MIT license. DeepSeek promoted the smaller Flash model to official status on July 31 and has now done the same for Pro.
