Chinese AI startup DeepSeek released DeepSeek-V4-Pro-0813 on August 12.

A brief note on its website said the model offered “significantly enhanced agent capabilities,” but that statement had been removed by August 13 afternoon as early developer feedback appeared mixed.

Independent benchmarks also produced mixed results.

On the Artificial Analysis Intelligence Index, the model scored 53, matching GLM-5.2 from Beijing-based AI company Zhipu AI, while trailing the Terra model in OpenAI’s latest GPT-5.6 series by four points and Moonshot AI’s Kimi K3 by seven.

San Francisco-based Vals AI ranked the model 12th on its index.

The company said DeepSeek-V4-Pro-0813 struggled with sandboxed terminal tasks and generating complex financial models in Microsoft Excel, although some researchers said it performed well in cybersecurity tests.

🔗 Source: South China Morning Post

Recent DeepSeek developments