This week, DeepSeek made an unexpected move by launching a public beta API for its updated V4-Flash model. The results, which I have carefully reviewed, overturn the usual model hierarchy: the "lightweight" version has not only caught up but, in many tests, surpassed the flagship V4-Pro-Preview. This is not just an incremental improvement but a serious restructuring of the competitive landscape in the agentic AI segment.

Key Changes in V4-Flash

The main focus of the update is enhancing agentic capabilities. The model now fully supports the Responses API format, allowing it to work autonomously: write and run code, search for information, and execute complex chains of actions. Moreover, V4-Flash has been adapted for integration with OpenAI's Codex—a strategically important step that gives developers the ability to use DeepSeek within familiar tools without changing their workflow.

The architecture and model size remain unchanged, but the V4-Flash-0731 version has undergone additional retraining. To switch to the new version, you simply need to specify the model name deepseek-v4-flash in the API settings—no other changes are required.

Test Results: The Numbers Speak for Themselves

The benchmark numbers are impressive. In the Terminal Bench 2.1 test (executing tasks in the command line), V4-Flash scored 82.7 points compared to 61.8 for the previous version. This is higher than China's GLM-5.2 (81.0) and nearly on par with the leader—Anthropic's Opus-4.8 (85.0).

The progress in coding is especially noticeable. In the NL2Repo test (converting a text description into a ready-made repository), the model scored 54.2 points versus 39.4 previously. But the most dramatic leap is in DeepSWE (solving real-world programming tasks): 54.4 points versus a meager 7.3 for the previous version. That is nearly a sevenfold increase.

The progress in cybersecurity is also impressive. In the Cybergym test (finding vulnerabilities in code), V4-Flash scored 76.7 points versus 38.7 for its predecessor and 52.7 for the older V4-Pro-Preview. Across all other agentic tests—Toolathlon, Agents' Last Exam, AutomationBench—the new version also confidently outperforms previous generations.

My Analysis: What This Means for the Market

This move by DeepSeek is not just an update but a signal of a paradigm shift. The company shows that a model's effectiveness is determined not by size but by the quality of training. For developers, this means access to performance on par with top-tier closed models at a significantly lower price. The full release of V4-Pro, expected later, now looks even more intriguing—if the "younger" model is this good, what will the flagship show? I am closely following developments, and this trend is definitely worth considering when choosing tools for agentic development.