DeepSeek pushed its flagship model to general availability on August 12, 2026, ending a preview that ran nearly four months. The GA build, tagged DeepSeek-V4-Pro-0813, is a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters per token and a 1 million token context window.

What This Enables

V4 Pro slots in as a high-reasoning backbone for agent and coding pipelines. You can point an existing OpenAI ChatCompletions or Anthropic Messages client at DeepSeek's endpoint and swap models with a one-line config change, because the API speaks both formats plus DeepSeek's own Responses API. Three reasoning modes (non-think, think-high, think-max) let you trade latency for depth on a per-call basis. Builders who prefer to self-host can pull the MIT-licensed weights from the Hugging Face model card and run the model on their own hardware instead of paying per token.

Why It Matters for Creators

A 1.6T flagship with a 1M-token window at $0.435 per million input tokens is one of the cheapest frontier-tier options for long-document and full-codebase work. For anyone building writing assistants, research tools, or coding agents, that pricing makes it practical to keep an entire manuscript or repository in context without truncation. The MIT license also lets you fine-tune and ship commercial derivatives, which most closed flagships forbid, though DeepSeek has signaled upcoming API price increases.

Key Details

Model: DeepSeek-V4-Pro-0813, 1.6T total and 49B active MoE, pre-trained on 32T-plus tokens

Context: 1M tokens in, up to 384K tokens out

Pricing: $0.435 per 1M input on a cache miss, $0.87 per 1M output, cache hits drop to $0.003625 per 1M

Benchmarks: 80.6% SWE-bench Verified, 93.5% LiveCodeBench, 90.1% GPQA Diamond

Access: API on OpenRouter and DeepSeek, plus MIT open weights on Hugging Face

The DeepSeek API docs confirm the deepseek-v4-pro endpoint now serves the 0813 build across ChatCompletions, Anthropic Messages, and Responses formats, with tool calling and JSON output supported.

What to Do Next

If you already run the open DeepSeek V4 Flash for agent workloads, benchmark V4 Pro against it on your own tasks before switching. DeepSeek says the re-trained Flash build beat the preview Pro on internal coding-agent suites, so Pro is not automatically the better default. Test the managed endpoint through DeepSeek if you want hosted inference, or self-host from the model card to control cost and data.