DeepSeek

The $5.6 Million Disruption

DeepSeek claimed they trained their V3 model for roughly $5.6 million. Whether that number includes everything is debatable, but the result isn't: a model that competes with GPT-4o at a price point that sent Nvidia stock tumbling 17% in a single day. The Hangzhou-based lab proved you don't need billions in compute to build world-class AI.

Technical Performance That Surprised Everyone

DeepSeek R1 is the standout. On math benchmarks like AIME 2024, it scores competitively with o1. Code generation in Python and common languages is strong — the model handles algorithmic problems and data structure questions at a level most developers would find genuinely useful. The reasoning chain is transparent: you can watch the model think through problems step by step.

DeepSeek V3 handles general conversation, writing, and analysis well. The output isn't as polished as Claude for nuanced writing, but it's surprisingly close to GPT-4o quality for most practical tasks. Chinese language support is obviously excellent, and English output has improved dramatically from earlier versions.

The Open-Source Angle

This matters enormously. DeepSeek releases model weights under permissive licenses. Companies can download, fine-tune, and deploy these models on their own infrastructure. For organizations that can't send data to external APIs — healthcare, defense, finance — self-hosted DeepSeek is a legitimate option. The community has already built quantized versions that run on consumer GPUs.

The China Question

Let's address this directly. DeepSeek is headquartered in Hangzhou, China. Chinese data privacy laws mean the government can theoretically compel access to user data. For casual use, this probably doesn't matter. For business use with sensitive data, it's a genuine concern. The self-hosted option sidesteps this entirely — download the model weights and run locally.

Several countries have restricted or investigated DeepSeek's data practices. Italy temporarily banned the app. The US government flagged potential security concerns. Individual users should evaluate their own risk tolerance.

Pricing That Changes the Economics

The web chatbot is free. No limits that feel restrictive for normal use. The API pricing is where things get wild: $0.14 per million input tokens and $0.28 per million output tokens for V3. Compare that to GPT-4o at roughly $2.50/$10 per million tokens. For companies running AI at scale, switching to DeepSeek's API can cut costs by 90% or more.