Deepseek V4 Pro has been released. Apart from its multimodal capabilities, it meets all my expectations for the model.

After DeepSeek V4 Flash was launched with stunning ultra-high cost-effectiveness, people are increasingly looking forward to what kind of performance ceiling the official version of DeepSeek V4 Pro will bring.

Tonight, DeepSeek's official documentation updated the model date: Deepseek-V4-Pro-0813, which means the official version of the model is here. It is already available in the official API for now.

V4 Pro adopts MoE design. According to the previously public architecture specifications, its total parameter scale is 1.6 trillion, with about 49 billion parameters activated per inference. There should be no changes in this official version.

The V4 Pro interface provides a 1 million-token context, a maximum output of 384,000 tokens, and supports the thinking toggle, tool calling, structured JSON, Responses API, and two compatible formats for OpenAI and Anthropic.

In other words, what DeepSeek has enhanced this time is not only how "deeply the model can think", but also significantly strengthened its Agent capabilities by optimizing the loading of large code repositories, understanding of massive amounts of data, and continuous tool calling to complete long-process tasks.

The benchmark testing session is still essential. As can be seen from the figure below, there have been earth-shaking changes compared to the preview version, and the official version is infinitely close to Fable 5.

The Terminal Bench 2.1 score of the V4 Pro preview version was only 72.1, while the official version directly jumps to 87.9, a total increase of 15.8 points in one go.

The DeepSWE performance is even more impressive, soaring from 12.8 to 62.7, nearly 4.9 times the original score. The preview version was basically "unable to complete tasks" in this benchmark, while the official version can now compete head-on with the world's first-tier code Agents.

Several other long-task evaluations also show substantial improvements instead of incremental minor upgrades.

AutomationBench rose from 12.8 to 31.8; DSBench-FullStack increased from 41.8 to 71.1; DSBench-Hard jumped from 31.1 to 67.2, more than doubling with extra additional gains; the cybersecurity offensive and defensive test CyberGym also rose from 52.7 to 83.3.

Next, let's look at the horizontal comparison with other models.

On Terminal Bench 2.1, V4 Pro scored 87.9, only 0.4 points behind Kimi K3's 88.3, and higher than Opus 4.8's 85.0 in DeepSeek's comparison chart.

In terms of high-difficulty reasoning, its score with HLE tools enabled is 60.0, second only to Fable 5's 63.0, and significantly higher than GLM-5.2's 54.7.

In terms of code engineering, NL2Repo scored 61.5, lower than Opus 4.8's 69.7, but has firmly secured a position in the first tier. DSBench-FullStack and DSBench-Hard reached 71.1 and 67.2 respectively, with no obvious shortcomings when facing complete engineering tasks and high-difficulty data science tasks.

When it comes to cybersecurity offensive and defensive tasks, V4 Pro scored 83.3 on CyberGym, even slightly higher than Fable 5's 83.1 and Opus 4.8's 78.3.

What is really noteworthy is that almost no capability of the official V4 Pro version falls out of the first tier.

It is not rare to rank first in a single item. Some models excel at coding, some at tool calling, and some have exceptionally high reasoning scores. The hard part is having no shortcomings at the same time in terminal operation, code engineering, tool calling, cybersecurity offense and defense, and long task execution.

Then the price is also quite competitive:

As of press time, the official price of DeepSeek Pro version per million tokens is: 0.025 yuan for cache-hit input, 3 yuan for cache-miss input, and 6 yuan for output.

It is indeed more expensive than the Flash version which charges 1 yuan for input and 2 yuan for output, but it is still surprisingly low when compared with other flagship Agent models.

For reference, the current API price of Kimi K3 is 3 USD for cache-miss input and 15 USD for output; Fable 5 even charges 10 USD for input and 50 USD for output. Even considering the currency difference, the price of DeepSeek is only a tiny fraction of that of overseas models of the same level.

Moreover, it provides a 1M context window and a maximum output of 384K tokens.

These two figures may not be very meaningful for ordinary chat scenarios, but they are rigid requirements for long-task Agents. Large code repositories, dozens of documents, long-running logs and continuous tool calls will quickly consume the context window. If the window is not large enough, the Agent can only keep compressing its memory, and eventually forget what it did earlier.

In addition, API compatibility has not been overlooked.

V4 Pro currently supports the OpenAI format, Anthropic format, Responses API, Tool Calls and JSON Output. The official team has provided access solutions for Agent tools such as Claude Code, OpenCode and OpenClaw. For existing workflows, the migration cost is much lower than re-adapting to a set of private protocols.

Of course, there are also trade-offs.

The maximum concurrent access limit for V4 Pro accounts is 500, while that for Flash is 2500. Its stronger capability and longer single-task duration mean that it is more suitable for complex tasks and high-value workflows, rather than being used to blindly undertake all high-frequency small requests.

There is another important reminder: DeepSeek has noted on its pricing page that it plans to raise the overall API price in the near future, with a significant expected increase.

So the current price is more like a window period, which cannot be directly regarded as a permanent commitment.

All in all, the official V4 Pro version has further lowered the industry threshold with its skyrocketing Agent benchmark scores, 1M context window and extremely low floor-level pricing.

The model invocation name remains unchanged, but what you get through the old API today is no longer the V4 Pro you used yesterday.

Recently, Jing Ge has several projects that plan to access V4 Pro, and run it through truly complex tasks in scenarios such as code Agents, data analysis or automated workflows.

To verify whether it only performs well in benchmarks, or is already fully capable of handling real work tasks.

This article is from the WeChat Official Account "Jingxuan AI", author: Jing Ge, published with authorization from 36Kr.