Key Highlights:
- DeepSeek has released DeepSeek-V3.2 and DeepSeek-V3.2-Speciale models that rivals OpenAI’s GPT-5 and Google’s latest Gemini 3 Pro in reasoning and agentic capabilities.
- The V3.2-Speciale hits “gold-level” scores on IMO, CMO, and ICPC benchmarks.
- The V3.2 model adds tool-integrated reasoning. While Speciale isn’t available in tools yet, it trades speed and token efficiency for raw performance.
Earlier this year, Chinese AI startup DeepSeek stormed the Silicon Valley with the release of the R1 model, which was an immediate rival to OpenAI’s then most advanced o1 reasoning model. One of R1’s standout features was its affordability.
Late last week, the company followed up with a high-precision math reasoning model called DeepSeek-Match-V2. Now, the Chinese AI startup has released DeepSeek‑V3.2 and DeepSeek‑V3.2‑Speciale.
DeepSeek’s new V3.2-Speciale model hits “gold-level” performance
Both are “reasoning-first” models, which have been designed for AI agents, complex thinking, tool use and complex reasoning. The company claims V3.2 offers what amounts to “daily-driver at GPT-5 level performance.” Whereas, the high-end Speciale variant “rivals Gemini 3 Pro.” DeepSeek-V3.2 is already live on its app, web interface and via API. For the uninitiated, it’s an upgrade over its earlier V3.2-Experimental model. The Speciale model, however, is now only available via a temporary API endpoint.
Here it’s important to note that DeepSeek claims the Speciale model has already hit “gold-level” performance on some popular benchmarks like the 2025 International Mathematical Olympiad (IMO), the China Mathematical Olympiad (CMO), and the ICPC World Finals. That’s enough to put the new model in the same category where Google’s and OpenAI’s high-end models are topping the list.
Under the hood, DeepSeek has expanded its agent-training approach with a massive synthetic dataset: 1,800+ environments and 85,000+ complex instructions. Meanwhile, V3.2 now supports “Thinking in Tool-Use,” meaning reasoning is baked directly into the model’s ability to call external tools. In the real-world, this makes it possible to mix deep reasoning with practical executions.
Comparison with GPT-5 and Gemini 3 Pro in benchmarks
So how does the new model stand against GPT-5 and Gemini 3 Pro? DeepSeek itself positions V3.2 as comparable to GPT-5 for day-to-day tasks. Meanwhile, Speciale aims directly at Gemini 3 Pro — especially on reasoning-heavy workloads.
On pure math benchmarks like AIME 2025, for instance, Speciale reportedly scores 96.0 — topping out GPT-5 High’s 94.6 and matching or slightly beating Gemini 3 Pro’s 95.0 in some tests. When it comes to SWE Verified, DeepSeek’s new model grabs 73.1, which is slightly lower than the Gemini 3 Pro (76.2).
Not to forget, Speciale’s reasoning power comes at the cost of token consumption and latency. Solving Codeforces–style problems reportedly required around 77,000 tokens with Speciale, which is far above Gemini’s roughly 22,000 tokens.
Also read: OpenAI Launches GPT-5.1 With Smarter Reasoning, Better Tone & More Personality
DeepSeek still wants to fine-tune the models
DeepSeek also acknowledges that the V3.2 model still lags commercial frontier models when it comes to knowledge breadth, token efficiency, and performance in the most complex tasks. However, the team has assured that it will bridge the knowledge gap with extensive pre-training. It’s also worth noting that, unlike V3.2, Special is not available in any tool yet.
All that said, DeepSeek is no longer just a niche open-source startup that is quietly competing from the sidelines, it is positioning as one of the top rivals that can square off against GPT-5 and Gemini 3 Pro in reasoning and math. We can’t write-off DeepSeek, as it might just become the go-to for anyone who wants GPT-5-like power without closed-door restrictions.
