DeepSeek updated its flagship model to DeepSeek-V4-Pro-0813 on 13 August, according to its own API documentation. The version string encodes the date, and the quick-start and pricing pages both reflect it.
What the update adds
The headline changes are support for the native Responses API format and tool calls on the flagship — both aimed at agentic use rather than raw capability. The model keeps a 1M-token context and 384K maximum output.
The pricing, which is the story
V4-Pro costs $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output tokens. The lighter deepseek-v4-flash is $0.14 and $0.28. Sitting alongside those figures is a warning that DeepSeek plans to raise API pricing "in the near future, with a significant increase expected" — a company that built its position on undercutting Western labs telling customers the discount is ending.
The documentation contradicts itself
DeepSeek's changelog has not caught up. The updates page still tops out at the 31 July V4-Flash entry and explicitly states the V4-Pro API is unchanged. The evidence for 0813 is the version string on the live pricing and quick-start pages, not a release note.
No benchmarks, so no capability claim
DeepSeek published no benchmark deltas and no release post for 0813. This is a point update to a serving model, and nothing in the documentation supports reading it as a new generation.
Why the price line lands hard
DeepSeek's per-token pricing has been the reference point that pushed inference costs down across the market. If it rises significantly, the anchor moves — and every buyer who built a cost model on the assumption of permanent Chinese price pressure has to rebuild it.