跪拜 Guibai
← Back to the summary

DeepSeek V4-Flash Ships Post-Training Overhaul That Doubles Agent Benchmarks

DeepSeek V4 finally saw new progress on the last day of July.

On July 31, DeepSeek's official update log showed that the DeepSeek-V4-Flash official API is now in public beta. The official team emphasized that this update is not just a simple version number change, but a significant post-training optimization targeting Agent capabilities. In multiple code, tool-calling, and automated task tests, the official V4-Flash version significantly outperformed the previous DeepSeek-V4-Pro-Preview.

Snipaste_2026-07-31_16-33-09.png

However, it's important to note that the scope of this release is quite limited: currently, only the V4-Flash API interface has been upgraded. The official DeepSeek-V4-Pro version has not yet been released, and the DeepSeek app and web version models remain unchanged for now.

In other words, ordinary users opening DeepSeek right now might not feel a noticeable difference. But for developers, AI Agent users, and programming tool users, this update deserves close attention.

1. Core Changes in the Official V4-Flash Version

According to the official description, DeepSeek-V4-Flash-0731 is consistent with the previous DeepSeek-V4-Flash-preview in terms of model structure and parameter scale. The main changes are concentrated in the post-training phase.

This means that DeepSeek did not redesign the model architecture or announce new parameter scales in this update. Instead, through further training and optimization, the model performs more stably when executing actual tasks.

This upgrade approach is somewhat different from the traditional concept of a "major version upgrade."

It is more like a capability reinforcement targeting real-world usage scenarios. The focus is not on making the model better at "talking" in chats, but on making it more adept at completing a series of complex tasks, such as:

This is also the biggest difference between an Agent model and a standard chat model.

Standard models prioritize answer quality, whereas Agent models must also handle environmental changes, tool-call failures, context management, task decomposition, and multi-turn execution. What truly determines an Agent's capability is often not whether it can answer a single question correctly, but whether it can complete a complex task from start to finish.

2. Why the Official Benchmark Scores Matter

This time, the official team published multiple Agent capability test results and explicitly stated that the official DeepSeek-V4-Flash's benchmark scores far exceed those of DeepSeek-V4-Pro-Preview.

HOiddvbbsAAddep.jfif

Among them, the Terminal Bench 2.1 score is 82.7, approaching Opus-4.8 in the same comparison table; Cybergym reached 76.7, Toolathlon Verified reached 70.3, and DeepSWE also reached 54.4.

More notably, compared to the preview version, the improvement in some projects is very significant.

For example:

If these tests were completed under the same or similar evaluation conditions, this suggests that V4-Flash's improvement is mainly reflected in complex task execution capabilities, not just text generation quality.

Of course, model benchmarks cannot be fully equated with daily experience. The task types, evaluation frameworks, prompts, and tool environments of different test sets will all affect the final results. Therefore, this data is more suitable as a reference for capability trends rather than an absolute ranking.

The official team also explained the test conditions: in the Code Agent tasks within public benchmarks, the official V4-Flash used the DeepSeek Harness minimalist mode, with the max gear, top_p=0.95, temperature=1.0. Among them, DSBench-FullStack and DSBench-Hard are DeepSeek's internal test sets, used to evaluate full-stack development and Coding Agent challenges respectively.

This means whether real users can ultimately achieve the same results depends on the subsequent toolchain, Harness framework, and specific application scenarios.

3. Native Support for Responses API, Adapted for Codex

Another very important change in this update is that the official V4-Flash natively supports the Responses API format and has been adapted for Codex.

The Responses API is an interface form increasingly adopted by Agent frameworks. Compared to traditional chat interfaces, it is more suitable for handling complex multi-turn tasks, tool calls, state management, and structured output.

For developers, this means integrating DeepSeek V4-Flash is no longer just about "calling it as a chat model," but allows for easier integration into programming assistants, automated workflows, and multi-tool Agents.

Especially for tools like Codex, which focus on code execution, file modification, and terminal operations, they typically require models with more stable tool-calling capabilities and clearer task planning abilities. DeepSeek's official mention of "targeted adaptation for Codex" this time indicates it is actively entering the more complex programming Agent ecosystem.

From the available interface information, V4-Flash supports OpenAI-compatible formats, as well as capabilities like tool calling, JSON output, and context continuation, making it overall more suitable for integration into various development tools.

This part of the change might not be directly perceivable by ordinary users, but for developers, it could be more important than improvements in chat ability.

4. Price Remains V4-Flash's Biggest Advantage

Snipaste_2026-07-31_16-32-00.png

Beyond capability, the most noticeable aspect of DeepSeek V4-Flash remains its price.

According to the currently public pricing information, V4-Flash's cache-hit input price is about 0.02 RMB per million Tokens, the cache-miss input price is 1 RMB per million Tokens, and the output price is 2 RMB per million Tokens.

In contrast, the corresponding prices for V4-Pro are higher, especially the output price, which is about 6 RMB per million Tokens.

This means V4-Flash could become the low-cost base model for a large number of developers building Agents.

Agent applications typically consume more Tokens than standard Q&A. A complex task might require multiple rounds of thinking, tool calls, log reading, code modification, and test feedback. If each execution relies on a high-priced model, the overall cost can quickly escalate.

Therefore, V4-Flash's cost-effectiveness is particularly noteworthy for the following scenarios:

Even if models like GPT-5.6 Luna have significantly reduced prices, DeepSeek V4-Flash still has the opportunity to continue attracting a large number of cost-sensitive developers with its lower unit price.

5. Why Ordinary Users Might "Not Feel" This Release

Although the officially announced results are impressive, this update will not immediately change the experience for all users.

The reason is simple: this upgrade currently only covers the DeepSeek-V4-Flash API.

The official team has clearly stated that the DeepSeek-V4-Pro API, as well as the app and web models, have not been changed.

Therefore, if you primarily use DeepSeek's service through the web or mobile app, you likely won't see significant changes in the short term. The first to truly experience the new capabilities are developers calling V4-Flash via the API, or users of related programming tools and Agent frameworks.

This also indicates that the focus of DeepSeek's release this time is not a product update for ordinary users, but an infrastructure upgrade for developers and the application ecosystem.

As for the most anticipated official V4-Pro version, the official statement currently is that it "will be released as soon as possible."

6. What Else Will the Pro Official Version Bring?

We currently cannot determine what specific changes the V4-Pro official version will have.

But judging from the update approach of V4-Flash this time, DeepSeek may continue to strengthen the model's capabilities in complex reasoning, code development, long context, and Agent execution.

V4-Flash handles lower-cost, higher-efficiency tasks, while V4-Pro might take on more complex, higher-difficulty reasoning and professional work. If a clear capability and price stratification forms between the two, DeepSeek's model product line will become more complete.

It is worth noting that DeepSeek V4's release pace has indeed been quite restrained. There was no large-scale preheating, nor consecutive days of intensive publicity. Instead, the API documentation was quietly updated on the last day of July.

From a timeline perspective, DeepSeek has at least temporarily kept the promise of a "July release." It's just that the Flash official version arrived first this time, and the Pro official version still requires more waiting.

Conclusion

Overall, the highlights of DeepSeek-V4-Flash-0731 can be summarized in three points:

First, the model architecture and scale have not changed, but through post-training optimization, Agent capabilities have shown significant improvement.

Second, it officially supports the Responses API and has been adapted for programming Agent tools like Codex.

Third, while maintaining an extremely low price, its code execution, tool-calling, and automated task capabilities have been further enhanced.

However, this is not a comprehensive upgrade for all users. The app and web versions remain unchanged for now, and the official V4-Pro version has not yet been released.

So, for ordinary users, this is more like a "technical development worth paying attention to"; for developers, it might mean a cheaper, more suitable model for building Agents has officially entered the usable stage.