DeepSeek launches V3.2 and V3.2 Speciale, reasoning-first AI models for agents

DeepSeek has officially launched its latest AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, designed as reasoning-first models optimized for intelligent agents. DeepSeek-V3.2, the official successor to the experimental V3.2-Exp, is now available across App, Web, and API, promising enhanced contextual understanding, multi-step problem-solving, and improved reliability for agent-based workflows. DeepSeek-V3.2-Speciale pushes the boundaries of reasoning capabilities further and is currently accessible API-only, focusing on complex analytical tasks and multi-step decision-making.

Compared to OpenAI’s latest models - GPT-5.1, GPT-4.1, and their o-series reasoning-specialist models, DeepSeek emphasizes agentic reasoning and structured problem-solving. OpenAI’s models, meanwhile, excel at broad versatility, multimodal inputs, large context lengths, and creative tasks. While OpenAI models provide more features and higher performance on diverse applications, DeepSeek’s reasoning-first models offer focused efficiency for tasks demanding deep logic and agent-style workflows.

DeepSeek’s pricing is also significantly lower than OpenAI’s flagship models. The V3.2 line reportedly charges $0.28 per million input tokens (cache-miss) and $0.42 per million output tokens, making high-volume usage highly cost-effective. By contrast, OpenAI’s advanced models cost around $1.25 per million input tokens and $10 per million output tokens, reflecting their broader capabilities and infrastructure. Such a low price point for DeepSeek raises questions about trade-offs in compute resources and performance, but it remains attractive for agentic, high-volume applications.

The hardware powering these models has become a topic of scrutiny. DeepSeek reportedly trained its large-scale models on Nvidia H800 GPUs, a restricted-export variant of the H100 designed to comply with U.S. export rules. H800 GPUs provide lower memory bandwidth and fewer tensor cores than full H100s, making them legal for China but limiting peak performance. Reports also suggest that for newer releases, DeepSeek may deploy models on domestic AI chips, including Huawei’s CANN accelerators, to circumvent restrictions on advanced Nvidia hardware.

Allegations and investigations suggest that some of these chips may have been obtained via black-market or grey-market channels, often routed through Singapore, Malaysia, and Thailand as intermediaries. Media reports claim over $1 billion worth of high-end Nvidia chips entered China through indirect channels despite U.S. bans, with allegations that DeepSeek sought to use shell companies and foreign data centers to access restricted hardware. However, there is no publicly verified evidence proving illegal acquisition, and the company maintains that it used lawfully obtained H800 GPUs.

DeepSeek’s V3.2 launch thus represents a major technical advancement in reasoning-first AI, while also highlighting the geopolitical and hardware constraints shaping the AI industry. Its combination of advanced reasoning, agent optimization, low pricing, and complex supply-chain realities makes it a notable entrant in the global AI landscape.