What it is
DeepSeek is a Chinese AI lab known for releasing strong open-weight models at low cost. Its chat app on the web, iPhone and Android is free and has no paid subscription.
The app runs DeepSeek V4, released April 24, 2026. V4-Flash powers the fast Instant mode and V4-Pro the Expert mode; the DeepThink toggle switches between quick answers and step-by-step reasoning. Web search and file uploads are included.
Developers pay separately for the API, which is among the cheapest from any major lab.
Our review
The model is the reason to care. V4-Pro, which became generally available on 13 August 2026, is a 1.6 trillion parameter mixture-of-experts model (49 billion active) with a 1 million token context, and the free app exposes it as Expert mode next to the faster Flash-based Instant mode. The weights are published on Hugging Face under the MIT licence, so the same model can be run on your own hardware, which is unusual at this level.
The privacy policy, last updated on 10 February 2026, is the problem for most work use. It says DeepSeek collects, processes and stores personal data in the People's Republic of China, keeps inputs for as long as you hold an account, and uses them to train its models. An opt-out exists, but the policy describes it as a right you exercise by emailing privacy@deepseek.com rather than an in-app switch. Several governments and companies have restricted the app on work devices, and it filters politically sensitive topics in line with Chinese rules.
For developers the API is a separate and more interesting proposition. At peak rates V4-Pro costs $1.32 per million input tokens and $3.96 per million output; Flash is $0.30 and $1.20. Off-peak, which covers all weekends, Chinese public holidays and weekday hours outside 01:00 to 04:00 and 06:00 to 10:00 UTC, is half that, and cache hits are cheaper still. The API speaks both the OpenAI and Anthropic formats. Teams that pinned the old deepseek-chat and deepseek-reasoner names should have migrated: they stopped working after 24 July 2026.
The app itself is plain next to ChatGPT or Gemini, with no image generation and few integrations, and DeepSeek publishes no daily cap for free use; heavy users report slowdowns rather than a clear limit. The useful split is between the hosted service, which inherits the Chinese data terms, and the open weights, which do not, and that split is what separates DeepSeek from Qwen Chat or Kimi as much as model quality does.
Where it earns its place
A backend engineer running a batch summarisation job over support tickets
Cut the per-token bill on a high-volume, non-sensitive workload
Point an OpenAI-compatible client at api.deepseek.com and schedule the job off-peak. Ten million input and two million output tokens a day on V4-Pro costs about $21 at peak and about $10.56 off-peak, before cache discounts; Pro's concurrency limit is 500 requests.
A platform team at a European bank that cannot send data to China
Use DeepSeek-level quality without the hosted service
Download the MIT-licensed V4 weights from Hugging Face and serve them in-house; nothing touches DeepSeek's servers. The 1.6T-parameter Pro model needs a multi-GPU cluster, so Flash (284B total, 13B active) is the realistic starting point.
A university student who will not pay $20 a month
Get step-by-step help with maths and code for free
Use Expert mode with DeepThink in the web or mobile app. Keep personal and coursework-identifying details out of prompts, since chats are kept for the life of the account and used for training.
Before you commit
- Setup
- Sign up at chat.deepseek.com or in the iOS and Android apps; API keys come from platform.deepseek.com and work with OpenAI or Anthropic SDKs by changing the base URL.
- Data and privacy
- Stored in the People's Republic of China; inputs used for training, with opt-out by email to privacy@deepseek.com; retained for as long as the account exists.
- Cost at scale
- The app is free. API at peak: V4-Flash $0.30 in and $1.20 out per million tokens, V4-Pro $1.32 and $3.96; off-peak halves both; search and fetch tool calls cost $0.75 per 1,000, with harness calls billed from 1 November 2026.
- Limits
- 1M-token context and 384K max output on both models; concurrency 2,500 for Flash and 500 for Pro; V4-Pro has no vision input; free app caps are not published.
What's good
- V4-Pro free in the app with a thinking mode
- MIT-licensed open weights you can host yourself
- API at $1.32/$3.96 per million tokens for V4-Pro, half that off-peak
- 1M-token context and OpenAI- and Anthropic-compatible API
What's not
- Data stored in China and used for training by default
- Training opt-out is by email, not a setting
- Censors politically sensitive topics
- No published usage limits for the free app
Best for
Budget-conscious users and developers who want a strong free model for non-sensitive tasks.
Alternatives
DeepSeek: common questions
- Is DeepSeek worth it?
- DeepSeek gives away a near-frontier model for nothing, and charges for it in data: your chats are stored in China and used for training unless you email to opt out. We rate it 3.9 out of 5.
- How much does DeepSeek cost?
- Chat app is free with no paid tier; usage caps are not published. API per million tokens at peak: V4-Flash $0.30 input, $1.20 output; V4-Pro $1.32 input, $3.96 output. Off-peak is half price.
- What are the downsides of DeepSeek?
- Data stored in China and used for training by default; training opt-out is by email, not a setting; censors politically sensitive topics; no published usage limits for the free app.
- Who is DeepSeek best for?
- Budget-conscious users and developers who want a strong free model for non-sensitive tasks.
- How much does DeepSeek cost for a team or in production?
- The app is free. API at peak: V4-Flash $0.30 in and $1.20 out per million tokens, V4-Pro $1.32 and $3.96; off-peak halves both; search and fetch tool calls cost $0.75 per 1,000, with harness calls billed from 1 November 2026.
- Is DeepSeek safe for company data?
- Stored in the People's Republic of China; inputs used for training, with opt-out by email to privacy@deepseek.com; retained for as long as the account exists.
- What are the best DeepSeek alternatives?
- The closest alternatives we have reviewed are Qwen Chat, Kimi and ChatGPT.
Answers checked Oct 22, 2026.