DeepSeek · current

DeepSeek V4 Pro

DeepSeek V4 Pro is listed on DeepSeek’s Models & Pricing page with API id deepseek-v4-pro. The model version on that page is DeepSeek-V4-Pro-0813. DeepSeek’s V4.1-Flash announce said they were phasing out V4-Pro and, for a window, routing Pro requests to Flash. A later docs note says API service for V4 Pro continues after 14 Sep 2026 with unchanged billing until further notice. Prefer live pricing and changelog wording.

Status: Current
Price: $0.66 in, $1.98 out (off-peak, cache miss)US dollars per million tokens (word pieces) in and out, unless noted

Training

Parameter count and training recipe for Pro-0813 are not published on the pricing page. The V4 family was open-sourced at Preview (Hugging Face collection). Treat Flash’s 552B MoE numbers as Flash-only unless a Pro card states otherwise.

Inference

Thinking and non-thinking modes (default thinking). Tool calls, JSON, the Responses API, and the Anthropic API are supported. Vision is not supported on Pro. FIM is non-thinking only.

Context

The documented context window is 1 million tokens. Max output is 384K tokens (same pricing table as Flash).

Methods

Same thinking effort controls as Flash (low, high, max). Exact RL recipe is not published on the pricing page.

Benchmarking

DeepSeek has not published a Pro score table on the pricing page cited here. The Flash announce claims Flash ahead of Pro on some vendor tests. Treat that as a company claim.

Hardware

DeepSeek has not published Pro serving silicon details on the pages cited here. Open weights for the V4 family: see the DeepSeek V4 Hugging Face collection from the Preview release.

Access and pricing

Per 1M tokens (pricing page):

	Off-peak	Peak
Input cache hit	$0.022	$0.044
Input cache miss	$0.66	$1.32
Output	$1.98	$3.96

Concurrency limit listed: 500. Same peak-hour definition as Flash.