DeepSeek Releases Flagship V4 Pro AI Model

DeepSeek has officially released the general-availability version of its flagship model, DeepSeek V4 Pro, under the build designation V4 Pro 0813. This release on August 12, 2026, marks the end of a nearly four-month preview phase. Available via OpenRouter and DeepSeek's own API, the model maintains its preview pricing of $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens. It features a one-million-token context window, a maximum output of 384,000 tokens, and a concurrency limit of 500. However, DeepSeek has warned that a significant price increase is planned for the near future.

The V4 Pro is a mixture-of-experts system containing 1.6 trillion total parameters, with 49 billion active per token. It utilizes Compressed Sparse Attention and Heavily Compressed Attention, which reduce single-token inference compute to 27 percent and the KV cache to 10 percent of the requirements of the previous V3.2 generation at a million-token context. Pre-trained on over 32 trillion tokens, the model was refined using on-policy distillation to consolidate domain-specific experts. While the April preview weights remain on Hugging Face, where they saw over 1.4 million downloads last month, DeepSeek has not yet released the updated 0813 weights.

In its maximum reasoning mode, labeled V4-Pro-Max, the model achieves notable scores on key benchmarks. It resolved 80.6 percent of issues on SWE-bench Verified, matching Gemini-3.1-Pro and falling just behind Claude Opus 4.6 at 80.8 percent. On Terminal Bench 2.0, it scored 67.9 percent accuracy, trailing GPT-5.4 at xHigh effort, which scored 75.1 percent. On Humanity's Last Exam, it posted a 37.7 percent pass@1, compared to Gemini-3.1-Pro at 44.4 percent.

For developers and practitioners, the production release of V4 Pro delivers a stable, highly capable foundation for agentic workloads like coding assistants and multi-step automation. The API supports three reasoning modes—non-thinking, high effort, and max effort—and integrates seamlessly with OpenAI ChatCompletions, Anthropic Messages, and DeepSeek's own Responses API, complete with tool calling and JSON output. This flexibility allows engineers to deploy advanced reasoning capabilities without changing their existing integration patterns, though they should prepare for upcoming API price adjustments.