After finishing the real-world test of the official version of DeepSeek V4 Pro, I found that the next "kill threshold" is likely hidden in its Agent.

Overnight, three blockbuster models were released almost simultaneously in a rare collision.

Grok 4.6's benchmark scores on paper are close to Fable 5, showing strong momentum to become one of the top three overseas model providers.

Qwen3.8 Max with a total parameter count of 2.4T also opened its weights as scheduled; on the DeepSeek side, the official version of DeepSeek V4 Pro (deepseek-v4-pro) was launched synchronously, with its API interface made available to the public.

After looking at the benchmark chart of the official V4 Pro version, I can only say that this round of improvement is clearly visible to the naked eye.

But these days, we have seen too many models that never lose in benchmark tests but fail to deliver satisfying real-world experience. Therefore, we immediately connected deepseek-v4-pro to Codex and Claude Code to test its actual performance.

Is the official V4 Pro version really easy to use?

Let's start with the text generation capability first.

I asked the official V4 Pro version to imitate Lu Xun and write a short essay of about 400 words, mocking the modern phenomenon that people would rather go hungry than skip taking photos of their meals to post on Moments, while requiring the work to have a proper sense of the era, irony and humor.

The result is as follows:

To be honest, at first glance, there is still a clear gap between the output and the real writing style of Mr. Lu Xun.

It even directly borrowed Lu Xun's most recognizable classic sentence pattern at the beginning, with very obvious traces of imitation. But at least the official V4 Pro version did not mechanically interpret the "Lu Xun style" as semi-classical Chinese, obscure words and archaic function words.

It can grasp the rhythm of the sentence patterns, and knows how to gradually guide the behavior of "taking photos before eating" from a daily activity to the pursuit of display desire, face and evaluation from onlookers.

More importantly, the obvious "AI assigned essay flavor" of the whole article has been greatly reduced. Though it is still not comparable to a literary giant, it at least does not have the awkward template splicing effect that makes people feel totally unimpressed.

If imitating Lu Xun is more or less like an "exam" scenario, the following business email test is closer to daily work scenarios.

I assumed that an internal mistake in the company caused a one-week delay of a customized system for a key client, and asked it to write an apology email that admits full responsibility, proposes compensation, and meanwhile will not shake the client's confidence in cooperation.

The official V4 Pro version did not use those common vague workplace expressions such as "multiple influencing factors" or "project pace adjustment", but directly admitted that "this delay is completely caused by our internal mistake", and then proposed three compensation measures including extended warranty, additional free training and exclusive operation and maintenance support.

The whole email was finished in about 28 seconds, and the main text is almost ready for use after only minor information modification.

Interestingly, after finishing the email, it added an extra explanation of why it wrote in that way, just like a student who just handed in the exam paper and insisted on explaining his problem-solving ideas to the teacher.

After finishing the writing capability test, I connected the official DeepSeek V4 Pro version to Codex.

DeepSeek has natively supported the OpenAI Responses API, and the official team also directly provided the Codex access solution. After configuration, Codex CLI, the desktop version of ChatGPT and the Codex plugin for VS Code can share the same set of configurations.

I also tested several programming tasks.

First, I assigned it a very typical "full-stack" front-end requirement.

The requirement is to build a classic macOS-style Web OS, with all codes packed into a single HTML file, and integrated with applications including text editor, Terminal, code editor, games, file manager, drawing tool and video editor.

Judging from the completion level alone, this should be the version with the richest functions among similar models I have tested before. Multiple applications have their basic interfaces and interactions implemented, and the system framework is also relatively complete, except that the aesthetic design is rather conventional.

Then I tested an instant camera case, which requires full simulation of the whole process including pressing the shutter, flash, photo ejection and 8 to 10 seconds of chemical development, with retro color cast, slight noise and the surface texture similar to Polaroid photos overlaid.

Judging from the final output, the transition between animations is also quite natural and smooth.

Another test case directly uses Three.js and WebGL, asking it to generate an interactive black hole accretion disk, where the starry sky, accretion disk and light bending effect all change accordingly when the mouse adjusts the observation angle.

Facing performance-sensitive scenarios like WebGL, V4 Pro still tends to be too aggressive in controlling the number of particles, real-time calculation, rendering overhead and animation complexity. Just after opening the web page, my M2 computer has already started to drop frames obviously.

The last test case is a sci-fi spaceship cockpit HUD, which needs to process dynamic crosshair, radar, heading, speed, energy, target locking and inertial offset after mouse movement at the same time. The official V4 Pro version finally implemented all the main interaction logics.

Generally speaking, if there is no official deployment bug, the official V4 Pro version has improved the completeness of handling complex tasks. However, after the whole experience, I always feel that there is a lack of stunning highlights, and even in some places, the gap between it and V4 Flash is not as huge as I expected.

But in Agent scenarios, if Flash can already complete 80% or even 90% of the tasks, and Pro is only responsible for a few high-difficulty nodes, then a truly efficient system will dynamically decide when to call Flash and when to call Pro, instead of running the most expensive model all the time.

In other words, the capability gradient between different models itself can become the cost optimization space for Agents.

DeepSeek's next winning card is hidden in Harness

In terms of specifications, the DeepSeek V4 series has extended the context window to 1M tokens, with a maximum output of 384K tokens.

Ordinary chat scenarios basically do not need such a large window, but Coding Agents often need to process dozens of files, historical modifications, tool return results and long-running task status at the same time. The longer the context window is, the more project materials the model can keep available.

The cost of this is also self-evident: the number of tokens will rise rapidly.

This also explains why the public paid so much attention when DeepSeek announced that it would raise API prices in the near future.

For the current official V4 Pro version, the input price for cache hits is 0.025 yuan per million tokens, the input price for cache misses is 3 yuan per million tokens, and the output price is 6 yuan per million tokens. The prices for V4 Flash are 0.02 yuan, 1 yuan and 2 yuan respectively.

The input and output prices of Pro are basically three times those of Flash, but they are still very cheap in the global market.

DeepSeek itself has also reminded in advance that the overall price of its API services is planned to be raised in the near future, and the increase is expected to be significant. Therefore, the current 3 yuan input price and 6 yuan output price are more like a phased preferential price.

Another notable point is that according to the existing model information, its main capabilities are still concentrated in text generation, reasoning, coding, tool calls and other directions, and there is no native input capability for images, videos and other modalities for the time being. In short, multimodality is not the priority of the official V4 Pro version at present.

In contrast, around the release of the official V4 Pro version, DeepSeek's actions in the Agent direction have become obviously frequent.

In the past two or three months, we have seen Cui Tianyi, who officially took charge of DeepSeek Harness department, actively recruiting talents on various platforms, showing extremely strong demand for top professionals.

Including an official WeChat official account named "DeepSeek Harness Team" has also been registered recently, whose certification subject is DeepSeek, and no content has been published on it yet.

Harness is a layer in the current Agent competition that is often overshadowed by the halo of large models.

The model is responsible for reasoning, while the entire external runtime framework determines how the Agent reads files, calls tools, manages context, retries after failures and executes continuously.

The actual experience of Claude Code and Codex is never determined only by the underlying model.

Especially after DeepSeek has both Flash and Pro in its product line, Harness has another very practical task: determining which tasks are worth running on Pro, and which tasks are sufficient to be handled by Flash. Once the model routing mechanism is well implemented, the Agent does not need to bear the cost of Pro all the time just for a few high-difficulty steps.

On the aspect of model cost advantage, we have already witnessed DeepSeek's aggressive pricing strategy. It does not need to rank first in every capability. As long as the model reaches the level of "being able to complete the vast majority of tasks", and the calling price is kept low enough, it is enough to occupy a solid market share.

Similarly, the most noteworthy part of the DeepSeek Harness team now is how to further transmit the existing model cost advantage of DeepSeek to the entire Agent system.

Let the model read context less repeatedly, make tool calls more efficient, avoid meaningless consumption in long tasks, reduce the number of tokens required for the same task, and then stack the price advantage of the underlying model itself.

At that time, even if DeepSeek does not necessarily develop the next market-dominating Claude Code by itself, as long as more and more Claude Code-like products are built on this system with extreme cost performance, it is also a victory that cannot be ignored.

This article is from the WeChat Official Account "APPSO", the author is APPSO that discovers tomorrow's products, published with authorization from 36Kr.