Descriptions:
DeepSeek has released two new open-source models: DeepSeek V3.2, a standard model available on the web app and API, and DeepSeek V3.2 Special, a reasoning-focused model available via API only. TheAIGRID breaks down what makes this release significant — particularly the fact that both models achieve benchmark performance comparable to GPT-5 and Claude 4.5 Sonnet while being fully open-source.
The video walks through benchmark results including Humanity’s Last Exam (25% for standard, 30% for Special), CodeForces, and LiveCodeBench, where DeepSeek Special edges out GPT-5 High. The host manually adds Claude Opus 4.5 data, finding it still leads in areas like SWE-Verified and Terminal Bench. A notable technical improvement is dramatically flattened cost scaling at long context — V3.2 handles 128k-token inputs at near-flat cost versus V3.1’s steep pricing curve, which the research paper attributes to selective attention over relevant tokens. Pricing comparisons suggest DeepSeek V3.2 runs approximately 30x cheaper than GPT-5 for equivalent tasks.
The host also flags suspicious identical benchmark values in DeepSeek’s own comparison tables and cautions that benchmark performance does not always translate to real-world usability, particularly for V3.2 Special whose extended reasoning time makes API costs considerably higher than they might initially appear.
📺 Source: TheAIGRID · Published December 02, 2025
🏷️ Format: News Analysis
