What DeepSeek is best at
- Anyone who wants frontier-class answers for $0
- Startups routing high-volume API calls on a budget
- Coders who want near-frontier ability without a subscription
Where it isn’t the answer
Sensitive topics and a polished consumer experience - the app is functional, not delightful, and it is a China-based service; judge your data accordingly.
Plans & pricing
The story so far
- 2024DeepSeek emerges from quant-fund research
- Jan 2025R1 shocks the field: open reasoning at a fraction of the price
- Apr 24, 2026V4-Pro and V4-Flash ship together under the MIT licence
- Aug 13, 2026V4-Pro goes generally available with a big jump in agent capability
- Sep 10, 2026V4.1-Flash arrives on a new architecture and beats V4-Pro on cost, speed and quality
Questions people actually ask
Is DeepSeek really free?
Chat is free. The API is priced at a level that rounds to free for most personal use - $0.15 per million input tokens off-peak, $0.30 at peak. Peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays; everything else, weekends included, is off-peak.
Which DeepSeek model should I use?
V4.1-Flash, which is what the `deepseek-flash` API name now serves. DeepSeek’s own tests put it ahead of the older and much larger V4-Pro on quality, cost and speed. V4-Pro is still available on request, and a V4.1-Pro has been signalled.
How does it stay this cheap?
The new architecture splits reading from writing: V4.1-Flash activates only 8 billion of its 552 billion parameters to take in your prompt and 16 billion to produce the answer. Most of the model sits idle on any given token.
Is DeepSeek safe to use?
For general questions and code, millions use it happily. It is a China-based service - for sensitive personal or business data, apply the same judgment you would to any overseas cloud.
How is it so cheap?
Aggressive efficiency research - mixture-of-experts designs and training innovations that cut compute cost dramatically, then pricing to win the market.