China’s leading AI research lab, DeepSeek, has launched its latest reasoning-focused AI models designed to compete directly with OpenAI’s GPT-5 High and Google’s Gemini 3 Pro. The lineup includes DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The standard V3.2 model succeeds the earlier V3.2-Experimental version and now powers the DeepSeek app.
The more advanced DeepSeek-V3.2-Speciale delivers exceptional performance and is available exclusively through API access. This enhanced model secures gold-tier scores in IMO, CMO, ICPC World Finals, and IOI 2025, matching the level achieved by specialized systems from OpenAI and Google. A version of Gemini 2.5 Deep Think also earned a Gold medal at IMO 2025.
Although significantly more capable, the Speciale model requires higher token usage and currently lacks tool integration. Its agentic performance has also seen major improvements compared to earlier versions.
Moving on to benchmarks, DeepSeek-V3.2-Speciale scores 96.0 on AIME 2025, outperforming GPT-5 High (94.6) and Gemini 3 Pro (95.0). However, in Humanity’s Last Exam (HLE), the Speciale model records 30.6, while Gemini 3 Pro leads with 37.7.
In SWE Verified, DeepSeek’s new model reaches 73.1, slightly below Gemini 3 Pro (76.2). Overall, DeepSeek’s latest release demonstrates true frontier-level capability, reinforcing China’s position as a powerful competitor in the AI landscape. With models like Alibaba’s Qwen 3, DeepSeek continues to push strong benchmark performances across the board.
