DeepSeek V4 Pro has moved from preview talk into a more concrete API shape. On August 13, the DeepSeek pricing page listed deepseek-v4-pro as DeepSeek-V4-Pro-0813, with unchanged OpenAI-compatible and Anthropic-compatible base URLs.
The useful point is not another model name. The document now puts the service limits, price gap and concurrency ceiling in one place: 1M-token context, 384K maximum output, JSON Output, Tool Calls, Responses API, Anthropic API, and a 500 concurrency limit for Pro.
The ceiling is for long work
“We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected.”
That warning makes the launch more practical than celebratory. A 384K output window can help with long-codebase analysis, document review, migration plans and agent workflows, but it also moves the bill into the generation phase.
Pro costs three times Flash
DeepSeek's Chinese price table lists V4 Pro at RMB 0.025 per 1M cached input tokens, RMB 3 for cache-miss input and RMB 6 for output. V4 Flash is RMB 0.02, RMB 1 and RMB 2. In plain terms, Pro's cache-miss input and output are three times Flash, while its concurrency limit is 500 rather than 2,500.
That split matters. Flash 0731 already had official agent benchmark numbers in the July 31 changelog, including Terminal Bench 2.1 at 82.7 and Cybergym at 76.7. The new Pro API version should not be read as automatic superiority on every workload.
The model is large, the use case should be narrow
The Hugging Face model card and OpenRouter page both keep the base scale clear: DeepSeek V4 Pro is a 1.6T-total-parameter MoE model with 49B activated parameters and a 1M-token context. That makes it a serious high-end service tier, not a casual default for short chats.
The sensible routing is therefore selective: use Pro for long-context repository understanding, difficult agent plans, cross-document reasoning and offline analysis where higher per-call cost is acceptable. Keep high-frequency summaries, simple completion and routine extraction on cheaper models unless tests say otherwise.
The next checks are also concrete: whether DeepSeek publishes a separate V4 Pro-0813 report, whether third-party benchmarks distinguish it from V4 Pro Preview and V4 Flash 0731, and how the promised pricing change affects the long-context cost advantage.
Sources: DeepSeek API models and pricing documentation, DeepSeek API changelog, Chao News, Hugging Face DeepSeek-V4-Pro model card, CocoLoop, OpenRouter; verified the DeepSeek-V4-Pro-0813 version, 1M context, 384K maximum output, RMB and USD pricing scopes, concurrency limits, V4 Flash comparison prices, 1.6T total parameters and 49B active parameters.