DeepSeek releases faster AI models that now match leading competitors on reasoning tasks

DeepSeek, a Chinese AI lab, unveiled two new AI models with architectural improvements that make them faster and more capable than their previous version (V3.2), and claim near-parity with frontier models (like OpenAI's o1 and Claude) on reasoning benchmarks (tests that measure an AI's ability to solve complex, multi-step problems).

The models require less computing power to run while delivering better performance—meaning organizations can deploy them on cheaper hardware and get answers faster, lowering the barrier to building AI applications for startups and smaller companies.

If DeepSeek's claims hold up, it challenges the assumption that only expensive, Western-made AI models (OpenAI, Google, Anthropic) can handle hard reasoning tasks, which could pressure those companies on pricing and accelerate the shift toward open-source or lower-cost AI alternatives in enterprise software.