Original title: (DeepSeek V4 Pro official release is now live)
Original author: Insight Beating
DeepSeek V4 Pro is officially here.
On the evening of August 12, DeepSeek’s official API documentation has updated the model version corresponding to deepseek-v4-pro to DeepSeek-V4-Pro-0813. Developers can now directly call the new version via the API. Pricing is 3 RMB per million input tokens and 6 RMB per million output tokens. For cached-hit input, the cost is only 0.025 RMB per million tokens.
V4 Pro-0813 continues to support a 1-million-token context window, with a maximum output of up to 384,000 tokens, and also supports both thinking and non-thinking modes. Capabilities such as JSON Output and Tool Calls are also available.
As of the evening of August 12, DeepSeek’s official update log still had not separately released an announcement for the official V4 Pro version. However, when DeepSeek updated the official V4 Flash version earlier, it had explicitly said that the official V4 Pro version would follow later. Now, the API backend model nodes have switched from the earlier version to DeepSeek-V4-Pro-0813.
Nearly four months have passed since V4 first appeared.
On April 24 this year, DeepSeek released V4 Preview and also rolled out V4 Pro and V4 Flash. V4 Pro is the flagship version of the entire lineup, with a total parameter count of 1.6T. Each token activates about 49B parameters, and it supports a 1-million-token context. From the very beginning, this V4 generation clearly strengthened coding, tool calling, and long-duration agent tasks.
On July 31, DeepSeek released V4-Flash-0731. Officially, this version keeps the original model architecture and size, but mainly involves additional post-training. The agent capabilities are significantly improved, and a Responses API is added natively. It also has been adapted for Codex. At the time, DeepSeek specifically emphasized that this update only concerns Flash; there were no changes to the Pro and the web-portal models. The V4 Pro official version was released later.
Now, the last missing puzzle piece on the Pro side has finally been filled in.
This generation of DeepSeek models is increasingly built around agents.
The release timing of the official V4 Pro just happened to coincide with the latest shift in the industry’s focus for the big-model product cycle.
When this year’s model vendors release new products, there is increasingly less room for pure comparisons of knowledge Q&A, math, and traditional benchmarks. Coding, computer use, tool use, long-duration task completion rates, and—most importantly—how much it costs to complete a single task are becoming the new core metrics.
When OpenAI released GPT-5.6 in July, it spent a lot of effort introducing agent and coding capabilities. GPT-5.6 Sol supports a new max reasoning effort and further offers an ultra mode where multiple agents complete complex tasks in parallel. The Responses API also begins to allow models to directly write programs to coordinate tools, handle intermediate results, and then decide the next step. OpenAI even made “how much work can be done for each dollar” one of the main selling points of this generation.
Anthropic’s approach is also pretty close.
Claude Sonnet 5, released at the end of June, mainly improved the same areas focused on coding, agents, computer use, and long-duration tasks. Anthropic provides different effort levels so developers can choose reasoning strength based on the task. On August 10, Anthropic also announced that the promotional pricing of Sonnet 5—2 USD per million input and 10 USD per million output—would be permanently retained.
What model companies are facing now is not just a question of who is smarter.
An agent may run continuously for dozens of minutes or even hours—repeatedly reading and writing code, searching for information, calling a browser, and using external tools. Once it truly enters a production environment, the number of model calls increases rapidly. Capability, speed, token consumption, and call costs start together to determine whether an agent product can actually run.
DeepSeek V4’s design also largely follows this direction.
A million-token context allows the model to read a much larger codebase, enterprise knowledge base, and task history in one go. Tool Calls are responsible for interacting with external environments. Compatibility with the Anthropic API—and the fact that Flash already supports the Responses API—are what lower the barrier for models to enter existing agent toolchains. When DeepSeek released V4 in April, it already opened both OpenAI ChatCompletions and Anthropic-format interfaces.
This is also the part where the official V4 Pro version is more worth关注 than yet another DeepSeek model on its own. DeepSeek is turning the model into foundational infrastructure that agents can directly call.
After the official release of the V4 Pro version, DeepSeek’s product lineup for this generation also became more clearly tiered.
At present, the input price for V4 Flash is 1 yuan per million tokens, and for Pro it is 3 yuan. Flash output is 2 yuan, while Pro output is 6 yuan. The prices differ by exactly three times. V4 Pro and Flash both support up to a million-token context, but Pro uses a larger activation-parameter scale, positioning it more for complex reasoning, coding, and more difficult agent tasks.
Flash handles lots of high-frequency, cost-sensitive tasks, while Pro is left for heavier work.
Model competition in 2025 still often revolves around “which flagship model tops the benchmark scores.” By 2026, models are increasingly becoming true software infrastructure: they need to be smart, and also have stable APIs, long context windows, tool calling, sufficient concurrency, and pricing that developers can pencil out.
After the official release of V4 Pro, DeepSeek’s product outline for this V4 generation is finally complete.
Original link