DeepSeek ships DeepSeek-V4-Pro-0813, taking its flagship model out of preview
DeepSeek has quietly moved its flagship language model out of preview status, updating its official API documentation to show the deepseek-v4-pro endpoint now running build DeepSeek-V4-Pro-0813. The change was made without an accompanying blog post or announcement — the only public record is the documentation update itself, first spotted by developers monitoring the company's API docs.
What's new
DeepSeek's models-and-pricing page confirms the update directly: "the deepseek-v4-pro model has been updated to DeepSeek-V4-Pro-0813," with the note that "the calling method remains unchanged — simply use deepseek-v4-flash or deepseek-v4-pro to access the latest version." No new model name or endpoint is required to pick up the change.
The specs carried over from preview:
- Context window: 1,000,000 tokens
- Maximum output: 384,000 tokens
- Concurrency limit: 500 requests for the Pro tier (versus 2,500 for the cheaper Flash tier)
- Pricing per 1M tokens: $0.435 for cache-miss input, $0.003625 for cache-hit input, $0.87 for output
Both models support JSON output, tool calling, DeepSeek's Responses API, and an Anthropic-compatible Messages format. DeepSeek's pricing page also carries a standing warning that it plans "to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected," with details still to come by "official notice."
Context
The Pro update follows the same path DeepSeek took with its smaller sibling: DeepSeek-V4-Flash graduated from preview to an official build, DeepSeek-V4-Flash-0731, on July 31. Both models originally shipped as downloadable open weights under an MIT license, and DeepSeek has since been iterating on the hosted API versions in parallel with the open releases, using dated build suffixes (0731, 0813) the same way it has versioned prior releases.
DeepSeek has built a reputation for shipping substantial model updates with minimal fanfare — often just a documentation diff rather than a press cycle — and the 0813 build fits that pattern. Because the API alias (deepseek-v4-pro) doesn't change, existing integrations pick up the new build automatically the next time they call the endpoint.
Why it matters
DeepSeek's V4 family has been central to the argument that open-weight, China-based labs can match a meaningful slice of frontier-model capability at a fraction of the price of US incumbents. Taking Pro out of preview status is DeepSeek's signal that the model is now considered production-stable rather than experimental — relevant for any team building agentic or coding tools on top of it, since preview builds typically carry looser stability guarantees than GA ones.
The pricing footnote is the other notable thread: DeepSeek is GA-ing the model at preview-era prices while simultaneously flagging that a "significant" increase is coming. That gives developers a narrow window to lock in current usage patterns before the cost structure shifts, and it raises the question of how much of DeepSeek's price advantage over Western labs was always meant to be a promotional rate for the preview period rather than a durable cost position.