Thought Leadership
DeepSeek has released V4-Pro-0813. The AI model shows weaknesses in benchmarks but impresses when it comes to detecting security vulnerabilities.
The Chinese AI startup DeepSeek made an updated version of its flagship model available on August 13, 2026, under the name DeepSeek-V4-Pro-0813. The update follows the V4-Flash model released in July. The company briefly stated on its website that the model offers “significantly improved agentic capabilities,” before the note was removed again.
In initial benchmark tests, the model showed mixed performance. On the Artificial Analysis Intelligence Index, it scored 53 points, putting it on par with Zhipu AI’s GLM-5.2, but behind models from OpenAI’s GPT-5.6 series and Moonshot AI’s Kimi K3. On the Vals Index for complex tasks, the application ranked 12th. According to analysts, weaknesses appeared particularly in handling tasks within isolated terminal environments and in creating complex financial models in Excel.
Strengths in cybersecurity and pricing
Despite the muted reception in general benchmarks, the model impressed in the area of IT security. According to analyses by the Belgian cybersecurity company Aikido Security, DeepSeek-V4-Pro-0813 outperformed all other tested models at detecting system vulnerabilities. Although precision was low in some cases, the model identified significantly more security flaws than competing systems such as Anthropic’s Claude Opus 5 or Alibaba’s Qwen 3.8.
Pricing for API use met with mixed reactions. Costs come to 44 US cents per million input tokens and 87 US cents per million output tokens. That makes the execution price slightly higher than for lightweight models such as GPT-5.6 Luna, but significantly cheaper than Kimi K3. DeepSeek has already announced future price adjustments due to high demand.
Planned expansions for autonomous agents
Alongside the model update, DeepSeek is preparing the introduction of a framework called DeepSeek Harness. The software is intended to allow large language models to act as autonomous AI agents, similar to existing developer tools.
Open-source developers have already been invited to take part in the beta testing phase for the new platform. Official technical documentation for the 0813 model has not yet been published.
(Editorial Team)
