이미지: X — 뉴스 앰프 화면 갈무리
Summary
- DeepSeek has published general availability (GA) pricing for DeepSeek-V4-Pro
- The model, versioned DeepSeek-V4-Pro-0813, offers a 1-million-token context window, priced at $0.435 per million cache-miss input tokens and $0.87 per million output tokens
- The concurrency limit is set at 500 requests, lower than the lightweight Flash model's 2,500
- 모델 버전
- DeepSeek-V4-Pro-0813
- 컨텍스트 길이
- 100만 토큰 (최대 출력 38.4만 토큰)
- 입력 가격(캐시 미스)
- 100만 토큐당 0.435달러
- 입력 가격(캐시 히트)
- 100만 토큰당 0.003625달러
- 출력 가격
- 100만 토큰당 0.87달러
- 동시 처리 한도
- 500건 (Flash는 2500건)
The numbers behind the pricing
DeepSeek's DeepSeek-V4-Pro has moved into general availability (GA), and its pricing has been made public. The model is versioned DeepSeek-V4-Pro-0813, with a context length of 1 million tokens and a maximum single output of 384,000 tokens. Pricing is set at $0.435 per million input tokens and $0.87 per million output tokens. However, the cache-hit input price — applied when the same prompt is called again — drops to $0.003625 per million tokens, about one-hundredth of the cache-miss rate. The concurrency limit is capped at 500 requests, lower than the lightweight Flash variant's 2,500. Both models support switching between thinking and non-thinking modes, and both are compatible with the OpenAI and Anthropic API formats. One news account described the pricing as "insane."
What's going on here
DeepSeek-V4-Pro first appeared on Hugging Face on August 9. That initial release used FP8 (8-bit floating point) precision, was released under the MIT license, and its model card described it as designed for million-token-scale context handling. What's being announced now is best understood as the finalized pricing that came with the model's transition to a commercial API service. The practice of pricing cache hits and cache misses separately is a structure already used by major API providers such as OpenAI — it works by reusing identical input, such as system prompts or repeated documents, without recomputing it, thereby cutting costs. DeepSeek has a track record of triggering price wars among competitors with aggressive pricing, and the V4-Pro rates appear to be an extension of that same strategy. A 1-million-token context is roughly enough to feed an entire book's worth of text into the model at once, making it especially useful for tasks like summarizing long documents or analyzing large codebases.
So what changes
The DeepSeek ecosystem is expanding rapidly around the V4 family. On August 11, Together AI began offering fine-tuning support for the lightweight DeepSeek V4 Flash 0731, while Nous Research is offering up to a 90% discount on V4 Flash through its own API gateway. With the Pro model's GA pricing set low, developers now have more room to test long-document-handling services at relatively low cost. That said, based on this source alone, it's unclear whether this pricing is a temporary promotion or a long-term policy.
