DeepSeek moved its flagship V4 Pro model to general availability on August 12, 2026, shipping the official 0813 build to its API and all downstream providers (OpenRouter, Vercel AI Gateway, Cloudflare, etc.). The model remained in preview since April 2026; this GA release marks the end of a four-month preview window. API pricing stays at $0.435 per million input tokens and $0.87 per million output tokens, with no code changes required for existing integrations.
The 0813 build is a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters, supporting 1 million token context and up to 384K output tokens. Updated internal weights aim to improve performance on reasoning, coding, and agentic tasks compared to the April preview. DeepSeek published benchmarks claiming competitive parity with closed-frontier models on software engineering (DeepSWE 62.7%, Terminal-Bench 87.9%), though independent verification on LMSYS Chatbot Arena and other platforms is still pending.
The architectural backbone remains unchanged: hybrid attention (Compressed Sparse Attention + Heavily Compressed Attention) reduces single-token inference to 27% and KV cache to 10% of V3.2’s cost at 1M-token context. DeepSeek supports three reasoning modes (non-thinking, high effort, max effort), tool calling, JSON output, and both OpenAI and Anthropic message formats. The model is available as MIT-licensed open weights on Hugging Face (~865GB, single-model deployment scale).
For agentic-pipeline architects: V4 Pro 0813 targets cost-sensitive text-only coding and reasoning workloads where 1.6T parameters and 1M context enable full-codebase analysis at ~1/60th the price of Claude Opus or GPT-5.6 Sol. Cache-hit rates on OpenRouter average 88-92%, dragging the effective input price to ~$0.05/M tokens. Watch for the announced price increase, which DeepSeek has flagged as coming soon; the arbitrage window on the current rates may close quickly.