Overview
- DeepSeek has listed Deepseek‑V4‑Pro‑0813 in its official API, making the formal V4 Pro model available to API users.
- V4 Pro offers a 1,000,000‑token context window and up to 384,000‑token output plus a ‘thinking’ switch, tool calls, and JSON/Responses API structured output.
- The model is reported as an MoE design with a stated ~1.6T total parameter budget and roughly 49B parameters activated per inference, features that support long, multi-step agent tasks.
- Published Pro pricing shows aggressive per‑million‑token rates (cache‑hit input 0.025 CNY, cache‑miss input 3 CNY, output 6 CNY) and a 500‑concurrency cap, and DeepSeek’s pricing page warns of an upcoming overall API price increase.
- Independent tests reported by coverage show large benchmark gains over the preview release across long‑task and code/agent evaluations, a result that positions V4 Pro as a low‑cost option for complex, long‑running automation and tool‑driven workflows.