🚨 In January 2025, a Chinese AI startup crashed Nvidia's stock by 17% in a single session. Training cost: $5.5M. GPT-4 cost: ~$100M. Same performance. Nobody saw it coming.
That startup was DeepSeek. We tested it for 3 weeks. Here's what we found.
The model that broke the industry
DeepSeek-R1 matched GPT-4o on every major benchmark at 1/20th the training cost. Within days:
- #1 most downloaded app on the US App Store
- Nvidia lost $600B in market cap
- Banned in 7+ countries before most people knew it existed
What actually impressed us
The visible Chain-of-Thought. R1 shows every step of its reasoning out loud — hypotheses, doubts, rejected paths. ChatGPT does something similar with o1. DeepSeek does it for free.
The code quality. 83.7% on SWE-bench Verified. We threw real tasks at it — REST API refactoring, Python debugging, React from specs. It matched Cursor and Copilot on most of them.
The API pricing. $0.30/M tokens vs $7.50/M for GPT-5. That's 25x cheaper for comparable performance.
The thing nobody talks about honestly
Your data goes to servers in China. Subject to Chinese law. A Wiz researcher found an unauthenticated database with 1M+ exposed records. It's banned on US government devices, at Microsoft, News Corp.
For sensitive data: don't use the cloud version. Run it locally with Ollama — same performance, your data never leaves your machine.
Our rating: 7.8/10
Technically impressive. Revolutionarily cheap. Real privacy trade-offs that can't be ignored.
→ Full review + API pricing breakdown + local deployment guide on Neuriflux
Originally published on Neuriflux — independent AI tools reviews, no bullshit.
Top comments (0)