DeepSeek Launches New AI Models Rivaling GPT-5

Hangzhou-based artificial intelligence (AI) company DeepSeek officially released two advanced models on Monday evening that demonstrate competitive performance against leading AI systems from OpenAI and Google. The new releases, named DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, represent significant progress in open-source artificial intelligence development.

DeepSeek-V3.2 delivers performance comparable to OpenAI’s GPT-5 across multiple reasoning benchmarks, according to company testing results. The model integrates reasoning capabilities directly into tool usage, allowing it to utilize search engines, calculators, and code executors while processing requests. This marks the first time DeepSeek has combined thinking functions with tool execution in a single model. The company describes DeepSeek-V3.2 as a daily driver model suitable for routine AI assistant tasks.

The second release, DeepSeek-V3.2-Speciale, focuses on advanced mathematical computations and extended reasoning tasks. Performance testing shows this high-compute variant matches Google’s Gemini-3.0-Pro on key metrics. DeepSeek-V3.2-Speciale achieved gold medal level results on the 2025 International Mathematical Olympiad and the International Olympiad in Informatics, demonstrating exceptional problem-solving abilities in technical domains.

Both models utilize the DeepSeek Sparse Attention mechanism, a technical breakthrough that substantially reduces computational complexity while maintaining model performance in long-context scenarios. The company implemented robust reinforcement learning protocols and scaled post-training computation to achieve these results. DeepSeek states its goal is to push inference capabilities of open-source models to their limits and explore the boundaries of model capabilities.

Founded in July 2023, DeepSeek focuses on research and development of large language models and multimodal AI technologies. The company previously disrupted the AI industry in January with a breakthrough model that demonstrated competitive performance despite lower development costs compared to Silicon Valley competitors. The latest releases build on momentum from recent innovations, including DeepSeekMath-V2, an open model with strong theorem-proving capabilities released last week.

The new models transition from experimental status to production-ready systems. DeepSeek previously released DeepSeek-V3.2-Exp in September as an experimental version. The company has now removed the experimental designation, signaling confidence in stability and reliability.

DeepSeek-V3.2 is available for free access through the company’s website, mobile application, and application programming interface. DeepSeek-V3.2-Speciale is currently accessible only through API access, with usage costs set at 42 cents per million tokens. This pricing represents a tenfold reduction compared to competing models from established AI providers.

Benchmark testing across reasoning and agentic capabilities shows DeepSeek models performing competitively against GPT-5, Claude-4.5-Sonnet, and Gemini-3.0-Pro. In reasoning assessments including AIME 2025 and HMMT 2025, DeepSeek-V3.2-Speciale achieved scores of 96.0 percent and 99.2 percent respectively. The model also demonstrated strong performance in agentic capabilities, scoring 73.1 percent on software engineering verification tasks and 80.3 percent on advanced benchmark tests.

The releases represent continued competition in the global AI model arena. OpenAI launched GPT-5 in August, describing it as the company’s smartest and fastest model. Google released Gemini-3.0-Pro in November as its latest AI system. DeepSeek’s ability to match performance of these proprietary models while maintaining open-source accessibility and lower costs presents an alternative approach to AI development.

DeepSeek has introduced new methodologies for training AI agents, which are software programs designed to act independently by interacting with environments, analyzing data, and making decisions without constant human oversight. The integration of reasoning with autonomous action execution in DeepSeek-V3.2 represents progress toward more capable AI systems that can handle complex tasks requiring both analytical thinking and practical implementation.

The company operates as an influential AI laboratory based in China, contributing to the country’s artificial intelligence research ecosystem. By releasing open-source models, DeepSeek enables developers and researchers worldwide to access advanced AI capabilities for specific applications. The text-focused nature of these models makes them suitable for language processing, mathematical problem-solving, and coding tasks, though they do not currently support image, video, or other multimedia processing.