DeepSeek V3.2 Matches GPT-5 Performance with 90% Lower Training Costs

While tech giants pour billions into computational power to train frontier AI models, China’s DeepSeek has achieved comparable results by working smarter, not harder. The DeepSeek V3.2 AI model matches OpenAI’s GPT-5 in reasoning benchmarks despite using ‘fewer total training FLOPs’ – a breakthrough that could reshape how the industry thinks about building advanced artificial intelligence.

For enterprises, the release demonstrates that frontier AI capabilities need not require frontier-scale computing budgets. The open-source availability of DeepSeek V3.2 lets organisations evaluate advanced reasoning and agentic capabilities while maintaining control over deployment architecture – a practical consideration as cost-efficiency becomes increasingly central to AI adoption strategies.

The Hangzhou-based laboratory released two versions on Monday: the base DeepSeek V3.2 and DeepSeek-V3.2-Speciale, with the latter achieving gold-medal performance on the 2025 International Mathematical Olympiad and International Olympiad in Informatics – benchmarks previously reached only by unreleased internal models from leading US AI companies.

The accomplishment is particularly significant given DeepSeek’s limited access to advanced semiconductor chips due to export restrictions.

Resource efficiency as a competitive advantage

DeepSeek’s achievement contradicts the prevailing industry assumption that frontier AI performance requires greatly scaling computational resources. The company attributes this efficiency to architectural innovations, particularly DeepSeek Sparse Attention (DSA), which substantially reduces computational complexity while preserving model performance.

The base DeepSeek V3.2 AI model achieved 93.1% accuracy on AIME 2025 mathematics problems and a Codeforces rating of 2386, placing it alongside GPT-5 in reasoning benchmarks.

The Speciale variant was even more successful, scoring 96.0% on the American Invitational Mathematics Examination (AIME) 2025, 99.2% on the Harvard-MIT Mathematics Tournament (HMMT) February 2025, and achieving gold-medal performance on both the 2025 International Mathematical Olympiad and International Olympiad in Informatics.

The results are particularly significant given DeepSeek’s limited access to the raft of tariffs and export restrictions affecting China. The technical report reveals that the company allocated a post-training computational budget exceeding 10% of pre-training costs – a substantial investment that enabled advanced abilities through reinforcement learning optimisation rather than brute-force scaling.

Technical innovation driving efficiency

The DSA mechanism represents a departure from traditional attention architectures. Instead of processing all tokens with equal computational intensity, DSA employs a “lightning indexer” and a fine-grained token selection mechanism that identifies and processes only the most relevant information for each query.

The approach reduces core attention complexity from O(L²) to O(Lk), where k represents the number of selected tokens – a fraction of the total sequence length L. During continued pre-training from the DeepSeek-V3.1-Terminus checkpoint, the company trained DSA in 943.7 billion tokens using 480 sequences of 128K tokens per training step.

The architecture also introduces context management tailored for tool-calling scenarios. Unlike previous reasoning models that discarded thinking content after each user message, the DeepSeek V3.2 AI model retains reasoning traces when only tool-related messages are appended, improving token efficiency in multi-turn agent workflows by eliminating redundant re-reasoning.

Enterprise applications and practical performance

For organisations evaluating AI implementation, DeepSeek’s approach offers concrete advantages beyond benchmark scores. On Terminal Bench 2.0, which evaluates coding workflow capabilities, DeepSeek V3.2 achieved 46.4% accuracy.

The model scored 73.1% on SWE-Verified, a software engineering problem-solving benchmark, and 70.2% on SWE Multilingual, demonstrating practical utility in development environments.

In agentic tasks requiring autonomous tool use and multi-step reasoning, the model showed significant improvements over previous open-source systems. The company developed a large-scale agentic task synthesis pipeline that generated over 1,800 distinct environments and 85,000 complex prompts, enabling the model to generalise reasoning strategies to unfamiliar tool-use scenarios.

DeepSeek has open-sourced the base V3.2 model on Hugging Face, letting enterprises implement and customise it without vendor dependencies. The Speciale variant remains accessible only through API due to higher token use requirements – a trade-off between maximum performance and deployment efficiency.

Industry implications and acknowledgement

The release has generated substantial discussion in the AI research community. Susan Zhang, principal research engineer at Google DeepMind, praised DeepSeek’s detailed technical documentation, specifically highlighting the company’s work stabilising models post-training and enhancing agentic capabilities.

The timing ahead of the Conference on Neural Information Processing Systems has amplified attention. Florian Brand, an expert on China’s open-source AI ecosystem attending NeurIPS in San Diego, noted the immediate reaction: “All the group chats today were full after DeepSeek’s announcement.”

Acknowledged limitations and development path

DeepSeek’s technical report addresses current gaps compared to frontier models. Token efficiency remains challenging – the DeepSeek V3.2 AI model typically requires longer generation trajectories to match the output quality of systems like Gemini 3 Pro. The company also acknowledges that the breadth of world knowledge lags behind leading proprietary models due to lower total training compute.

Future development priorities include scaling pre-training computational resources to expand world knowledge, optimising reasoning chain efficiency to improve token use, and refining the foundation architecture for complex problem-solving tasks.

See also: AI business reality – what enterprise leaders need to know

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.