DeepSeek

1:31 DeepSeek is a Hangzhou AI lab that offers a free chatbot at chat.deepseek.com and in mobile apps, and publishes its models as open weights under the MIT license. Its current models are DeepSeek-V4-Pro, generally available since Aug 13, 2026, and DeepSeek-V4.1-Flash-1017, released Oct 17, 2026, both with a 1-million-token context. API prices start at $0.15 per million input tokens and $0.60 per million output tokens off-peak (as of Nov 6, 2026).

DeepSeek is the pick when cost matters most or when you want strong models you can run on your own hardware. Its weak point is data: DeepSeek's privacy policy says it stores personal data in the People's Republic of China, which led Italy's privacy regulator to block it there. For sensitive work, a self-hosted copy of the open weights or a US or EU assistant such as ChatGPT, Claude or Le Chat is safer. See Best AI Chatbot.

Quick facts

Is DeepSeek free?

Yes. The DeepSeek chat app on the web, iOS and Android is free, with an Instant mode and an Expert mode that runs V4-Pro, per DeepSeek's Aug 13, 2026 release note. The API is pay as you go, and the open weights can be downloaded at no cost from Hugging Face.

API prices

3:59 Prices per million tokens. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; everything else, including weekends, is off-peak at half price. Source: DeepSeek Models and Pricing, read Nov 6, 2026. DeepSeek also sells web Search and Fetch APIs at $0.75 per 1,000 calls. In an Oct 20, 2026 video, Simon Høiberg reported that half a day of his agent work on V4.1-Flash-1017, run through OpenRouter, used 10.7 million tokens and cost $0.76.

Models

DeepSeek-V4 arrived as a preview on Apr 24, 2026, open-sourced from day one, with 1 million tokens of context as the default in all of DeepSeek's services. V4-Pro became generally available on Aug 13, 2026 with adjustable reasoning effort and support for OpenAI's Responses API, so it plugs into Codex and Claude Code.

1:46 V4.1-Flash-1017 (Oct 17, 2026) is the newest, which DeepSeek says scores higher than the previous Flash while writing shorter answers. Earlier models include V3 (Dec 26, 2024), R1 (Jan 20, 2025) and V3.2 (Dec 1, 2025), all on DeepSeek's news page.

History: the Jan 2025 shock

DeepSeek released its R1 reasoning model on Jan 20, 2025, free and open. A week later, on Jan 27, 2025, Nvidia lost 17% of its value, close to $600 billion, the biggest one-day loss in US market history, as investors weighed DeepSeek's claim that its earlier model took two months and under $6 million to build on Nvidia H800 chips.

Is DeepSeek safe?

DeepSeek's privacy policy (updated Feb 10, 2026) says it processes and stores personal data in China and uses what you type to train its models; you can opt out of training only by emailing privacy@deepseek.com, with no switch in the app. Italy's data protection authority ordered DeepSeek to stop processing Italian users' data on Jan 30, 2025 and opened an investigation, after DeepSeek said EU law did not apply to it. The open weights are a different matter: run on your own computer or a cloud you choose, nothing goes to DeepSeek. See Run AI Locally and AI Privacy Compared.

DeepSeek vs the alternatives

Qwen, Kimi and GLM are the other leading Chinese open-weight models; Moonshot's Kimi K3.1 Pro and Zhipu's GLM-5.3 compare themselves with DeepSeek V4 in their own benchmark tables. Le Chat (Vibe) is the European open-weight choice, with data handled under EU law. ChatGPT and Claude lead on top-end ability and keep data outside China, at higher prices. Pick DeepSeek for the lowest API cost or for strong open weights to host yourself.

Frequently asked questions

Is DeepSeek free? Yes. DeepSeek's chat app is free on the web, iOS and Android, and its V4 model weights are free to download under the MIT license; only the API is paid.

What is the latest DeepSeek model? DeepSeek-V4.1-Flash-1017, released Oct 17, 2026, alongside DeepSeek-V4-Pro, generally available since Aug 13, 2026.

Is DeepSeek safe to use? DeepSeek's app stores your data in China and trains on it unless you email to opt out, and Italy ordered it to stop processing Italian users' data in Jan 2025; running the open weights yourself avoids sending data to DeepSeek.

How much does the DeepSeek API cost? DeepSeek's Flash model costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, double at weekday peak hours, as of Nov 6, 2026.

Is DeepSeek open source? DeepSeek publishes its model weights under the MIT license on Hugging Face, so anyone may use, modify and sell them.

Sources

• Models and Pricing, DeepSeek API Docs, read Nov 6, 2026
• DeepSeek-V4.1-Flash-1017 Release, DeepSeek, Oct 17, 2026
• DeepSeek-V4-Pro GA Release, DeepSeek, Aug 13, 2026
• DeepSeek V4 Preview Release, DeepSeek, Apr 24, 2026
• DeepSeek-R1 Release, DeepSeek, Jan 20, 2025
• DeepSeek APP, DeepSeek, Jan 15, 2025
• DeepSeek-V4-Pro, Hugging Face, read Nov 6, 2026
• The new DeepSeek model is insane (best value for money right now), Simon Høiberg, Oct 20, 2026 (video)
• How To Use DeepSeek For Beginners, corbin, Jan 29, 2025 (video)
• DeepSeek Privacy Policy, DeepSeek, Feb 10, 2026
• Il Garante privacy blocca DeepSeek, Garante per la protezione dei dati personali, Jan 30, 2025
• Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history, CNBC, Jan 27, 2025