Chinese artificial intelligence startup DeepSeek recently launched two new powerful AI models, DeepSeek V3.2 and DeepSeek V3.2-Speciale. The company claims these models can challenge the performance of leading systems like Google's Gemini 3 Pro and OpenAI's GPT-5. The release marks a significant move by the Hangzhou-based firm in the competitive global AI landscape.
New Models Boast Top Reasoning Skills
DeepSeek V3.2-Speciale, the more advanced of the two new models, reportedly surpasses Gemini 3.0 Pro and GPT-5 High in several key benchmarks. DeepSeek stated its V3.2 model also matches or exceeds GPT-5 in reasoning tasks and performs "on par" with Gemini 3.0 Pro. These claims position DeepSeek as a serious contender against established American AI giants.
The V3.2-Speciale model achieved "gold-level results" in prestigious academic competitions, including the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). It also scored 99.2% on the Harvard-MIT Math Tournament and 73% on software bug fixing tasks. In coding, DeepSeek says V3.2 Speciale leads the CodeForces benchmarks.
Open-Source Approach and Cost Efficiency
A key aspect of DeepSeek's strategy is its commitment to open-source models. Both DeepSeek V3.2 and V3.2-Speciale are available under an MIT license, and their weights can be downloaded on Hugging Face. This open approach allows other companies and developers to utilize DeepSeek's advanced models freely.
DeepSeek also emphasizes cost efficiency in its AI development. The company claims the V3.2 model's pricing is significantly lower than its rivals. DeepSeek charges $0.28 for input and $0.42 for output per 1 million tokens. This contrasts sharply with Gemini 3 Pro's estimated rates of $2 for input and $12 for output, and GPT-5.1's $1.25 input and $10 output for the same token volume.
DeepSeek's Background and Previous Success
DeepSeek was founded in July 2023 by Liang Wenfeng, who also co-founded the Chinese hedge fund High-Flyer. High-Flyer provides funding for DeepSeek, allowing the AI firm to operate without external investors. Liang Wenfeng's personal fortune, largely based on his stake in DeepSeek, is estimated at $11.5 billion, with the company itself valued at approximately $15 billion as of November 2025.
The company gained attention earlier this year for developing high-quality AI models at a fraction of the cost of its competitors. For example, DeepSeek's V3 model reportedly cost $6 million to train, while OpenAI's GPT-4 cost an estimated $100 million. This cost-effective development has allowed DeepSeek to disrupt the AI industry.
DeepSeek-V2, an earlier general-purpose model launched in May 2024, utilized a "Mixture-of-Experts" (MoE) architecture. This approach helped reduce training costs by 42.5% and boosted generation throughput by 5.76 times compared to its predecessor, DeepSeek 67B. DeepSeek-V2 was trained on a vast corpus of 8.1 trillion tokens.
The company also developed DeepSeek-Coder-V2, an open-source model specifically for coding tasks. Released in July 2024, this model supports over 338 programming languages and handles context lengths up to 128,000 tokens. It has shown performance comparable to GPT-4 on code-specific tasks.
Impact on the Global AI Race
The release of DeepSeek's new models intensifies the global competition in artificial intelligence. Their strong performance and open-source availability challenge the dominance of U.S.-based AI firms like OpenAI and Google. This move highlights China's growing capabilities in advanced AI research and development.
Industry observers note that DeepSeek's ability to produce highly capable models at lower costs could put pressure on the valuations and pricing strategies of larger AI companies. OpenAI CEO Sam Altman previously acknowledged DeepSeek's models as "clearly great" and a reminder of the need for competition.
DeepSeek's strategy aims to make advanced AI more accessible and affordable, potentially reshaping how companies and developers interact with large language models. The company continues to focus on reasoning-first AI systems designed for agent workflows, pushing the boundaries of what AI can achieve.
