What DeepSeek is.
DeepSeek is a Chinese artificial intelligence company, based in Hangzhou, that builds language models: the same kind of technology behind ChatGPT, Claude or Mistral. It stands out through two strong choices:
- very low prices on its API;
- open models: the weights are published under the MIT license, so you can download them and run them yourself.
You can use it in three ways: the assistant (website and mobile app, free), the pay-as-you-go API to plug it into your tools, or the open models installed on your own servers.
The September 2026 models.
As of September 27, 2026, the DeepSeek API offers two models:
| DeepSeek-V4.1-Flash | DeepSeek-V4-Pro | |
|---|---|---|
| API model name | deepseek-flash |
deepseek-v4-pro |
| Positioning | Fast and low-cost | The most capable |
| Context | 1 million tokens | 1 million tokens |
| Maximum output | 384,000 tokens | 384,000 tokens |
| Images | Yes | No |
| Tool calls, JSON output | Yes | Yes |
| Thinking mode | Yes, on by default | Yes, on by default |
DeepSeek presents V4.1-Flash as an across-the-board improvement on text and agent tasks, with native image understanding. The legacy names (deepseek-v4-flash) are still accepted, but requests are now served by V4.1-Flash.
A practical detail for technical teams: the API accepts both the OpenAI and the Anthropic formats. A tool already wired to either connects to DeepSeek by changing the base URL and the key. The documentation lists Claude Code, GitHub Copilot and OpenCode among compatible tools.
Real API pricing.
DeepSeek bills per million tokens, input and output counted separately. One particularity: two rate cards depending on the time of day. Here are the prices shown on September 27, 2026, in US dollars:
| Per million tokens | V4.1-Flash, peak | V4.1-Flash, off-peak | V4-Pro, peak | V4-Pro, off-peak |
|---|---|---|---|---|
| Input (cache hit) | $0.006 | $0.003 | $0.044 | $0.022 |
| Input (cache miss) | $0.30 | $0.15 | $1.32 | $0.66 |
| Output | $1.20 | $0.60 | $3.96 | $1.98 |
Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. In US Eastern Daylight Time, that is 9 pm to midnight (Sunday to Thursday evenings) and 2 am to 6 am (Monday to Friday). Everything else, weekends included, is off-peak, at half price.
For scale: processing one million input tokens and generating one million output tokens with V4.1-Flash, off-peak, costs $0.75. For comparison, Mistral lists Mistral Large at $0.5 input and $1.5 output per million tokens. Two tips to pay even less: run bulk jobs off-peak, and reuse the same prompt prefixes to benefit from the cache.
Where your data goes.
This is what matters most for a business. DeepSeek's privacy policy states that the company collects, processes and stores personal data in the People's Republic of China. It has appointed a representative in the European Union for data subject requests, and specifies that its services are not designed to process sensitive personal data.
Our recommendation is simple:
- To test, draft, explore without confidential data: the free assistant and the API are fine.
- For customer, HR, financial or strategic data: do not go through the public app or API.
The good news is that you can keep costs low without exposing your data: the models are open.
Using DeepSeek without exposing your data.
DeepSeek publishes the weights of DeepSeek-V4.1-Flash and DeepSeek-V4-Pro on Hugging Face, under the MIT license, which allows commercial use. In practice, you can run them:
- on a rented GPU server in the region of your choice, so data never leaves it;
- through an inference provider hosting in your region, if it offers these models;
- on your own machines, if you already have the infrastructure.
The trade-off: these models are large and need suitable GPUs. The math is case by case, depending on the volume to process. For occasional use, a European provider like Mistral can be simpler: see Mistral's Le Chat Is Now Vibe.
DeepSeek in business: the right uses.
DeepSeek is especially attractive when volume is high and data is not sensitive:
- Bulk processing: classify, summarize or extract information from thousands of public or non-confidential documents, off-peak.
- Agents and automations: the fast, cheap model of an agent that chains many calls, with tool calls and JSON output.
- Development: a low-cost coding model in compatible tools, for projects without confidential code.
In a well-designed system, DeepSeek is just one of the possible models: each task goes to the model with the best trade-off between cost, quality and confidentiality. That is the approach laid out in Which AI to Choose in 2026, and the one we apply in our custom AI agents.
DeepSeek: what people ask us.
What is DeepSeek?
A Chinese artificial intelligence company, based in Hangzhou, that builds language models. It offers a free assistant (website and mobile app), a pay-as-you-go API, and publishes its model weights under the MIT license.
Is DeepSeek free?
The assistant is: the app is free on the App Store, with no in-app purchase listed, and the chat is available on the web. The API is paid per million tokens, deducted from a balance you top up in advance.
How much does the DeepSeek API cost?
As of September 27, 2026, DeepSeek-V4.1-Flash costs $0.30 per million input tokens and $1.20 per million output tokens at peak hours; DeepSeek-V4-Pro costs $1.32 and $3.96. Off-peak, prices are halved, and cached input costs far less.
What is the difference between DeepSeek V4 Flash and V4 Pro?
V4.1-Flash is the fast, low-cost model, with image understanding. V4-Pro is the most capable model, without vision, three to four times more expensive. Both accept 1 million tokens of context, tool calls, and a thinking mode that is on by default.
Is DeepSeek safe for business data?
According to its privacy policy, DeepSeek collects, processes and stores personal data in the People's Republic of China. For customer or sensitive data, we recommend not using the public app or API, and instead hosting the open models on infrastructure you control, or picking another provider.
Can I run DeepSeek on my own server?
Yes. DeepSeek publishes the weights of DeepSeek-V4.1-Flash and DeepSeek-V4-Pro on Hugging Face under the MIT license, which allows commercial use. You do need GPU servers sized for these large models.