DeepSeek V4: we tested the cheap open AI model, and checked where your data goes

DeepSeek is a Chinese AI company that gives its models away. Anyone can download DeepSeek V4 for free. And if you pay to use it online, it costs a fraction of what OpenAI, Anthropic or Google charge. We tested both DeepSeek V4 models on our own coding tasks. They passed the same secret tests as the expensive models. All twelve tries together cost us 7.4 cents. But there is one thing to know first: where your text goes.

What is DeepSeek V4?

DeepSeek V4 is a family of AI models from the Chinese company DeepSeek. There are two you can use now: V4.1 Flash and V4 Pro. V4.1 Flash is the newer one. It appeared on Hugging Face, the big website for sharing AI models, on 10 September 2026.

The special thing: DeepSeek puts the whole model online for free, under the MIT licence. That licence lets almost anyone use, change and even sell things built with it. The top models from OpenAI, Anthropic and Google cannot be downloaded at all.

The model is huge: DeepSeek says V4.1 Flash has 552 billion parameters at its core. Parameters are the numbers a model learns during training. But for each word it reads, only a small part of the model works, which is why it is cheap to run. It can read up to one million tokens at once, and it understands both text and images, according to DeepSeek. It has been downloaded more than a million times, Hugging Face shows. The V4 Pro model page is on the same site.

How much does it cost?

If you do not want to run the model yourself, you can pay DeepSeek to use it online, through its paid connection for developers. DeepSeek's price page lists, per million tokens, at the normal rate:

"Repeated text" is text the model has seen before in the same work, which it can reuse much more cheaply.

Compare that with the other models we tested. GPT-6.1 Sol and Claude Sonnet 5.5 cost $2 to read and $10 to write. Grok 4.7 costs $2 and $6. Even Google's cheaper Gemini 3.8 Flash costs $0.75 and $3.75, as we listed in our Gemini 4 Argon article. So V4.1 Flash writes for about one eighth of the price of Sol or Sonnet.

Half price outside busy hours

DeepSeek also has a cheaper rate at quiet times of the day. "Off-peak rates are half of the peak rates," its price page says.

The busy hours are on weekdays: 01:00 to 04:00 and 06:00 to 10:00, in UTC time. In the Netherlands in summer, that is 03:00 to 06:00 and 08:00 to 12:00. All other hours are half price, and so are weekends and Chinese public holidays.

So if you have a big job that is not urgent, starting it in the afternoon or at the weekend halves the bill. For most people the amounts are tiny either way, as our test shows.

Our test: cheap and just as good on our tasks

We gave both models two coding tasks from our earlier tests. In the first, they had to fix a bug in code that reads times like "1h30m". In the second, they had to build a Dutch invoice calculation from a description. Each model did each task three times. We checked every answer with our secret tests: tests we wrote ourselves and only ran after the AI had finished.

Both models passed exactly the same secret tests as every other model we tested, including GPT-6.1 Sol, Claude Sonnet 5.5 and Grok 4.7. And they did it for almost nothing. For comparison: the same six tries cost 25 cents with Grok 4.7.

There was a surprise, though. The "cheap" V4.1 Flash was not the cheapest per job. It thought much longer before answering: about 105,000 tokens of writing over six tries, almost all of it hidden thinking that you pay for but do not see. V4 Pro needed about 25,000. So Flash cost 5.1 cents and Pro only 2.3 cents. The lesson from our Grok test holds here too: the price per token is only half the story.

We ran the test through OpenRouter, a service that gives access to many AI models with one account. It does not always send your request to DeepSeek itself. In our test, V4.1 Flash was answered by CoreWeave and InferenceNet, and V4 Pro by StreamLake and Relace: companies that run DeepSeek's open model. Their prices can differ a little from DeepSeek's own list.

One bug-fix test failed for every model we have ever tested, including these two. It checks a rule our task did not mention, so we do not count it against them. You can compare with Grok 4.7 and Codex and Claude Code, which we tested the same way.

Where does your data go?

This is the part to read before you paste anything private into DeepSeek.

DeepSeek's privacy policy is clear: "To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China." The company behind the service is Hangzhou DeepSeek Artificial Intelligence, registered in China. The policy also says DeepSeek may use your data to train its models, and that you have the right to opt out of that.

For people in Europe, the policy has an extra section with rights under European privacy law, such as asking which data DeepSeek has about you. It does not change where the data is stored.

So for anything private or business-related, think twice before using DeepSeek's own app or website.

Three ways to use DeepSeek V4

Because the model is free to download, you are not stuck with DeepSeek's own service. There are three ways:

• DeepSeek's own app or service. Cheapest and simplest. Your data is stored in China, says the privacy policy.
• Another company that runs the model. Several companies host DeepSeek V4, such as the ones that answered in our test. Then that company's privacy rules apply. Check them before you use it.
• Your own computer. Free, and your data stays with you. But a model this size needs very expensive hardware. Smaller, squeezed versions made by others exist; Hugging Face lists about a hundred.

Our advice: DeepSeek V4 is a very cheap and capable option for work that is not secret, like learning to code or testing ideas. For private or customer data, pick a host whose rules you trust, or another model.

How we checked

On 6 October 2026 we read DeepSeek's price page, its privacy policy and the model pages for V4.1 Flash and V4 Pro on Hugging Face. The screenshots are our own captures of those pages, cropped only. The prices of other models come from their makers' own pages, read between 4 and 6 October.

For the test we used the same two tasks and the same secret tests as in our earlier tests, through OpenRouter, three tries per model, with default settings. We wrote down the time, the cost OpenRouter reported and which company answered. The owner of this site paid for the requests. The charts were made by us. None of the images was made by AI.

This was a small test with two clear tasks. It says nothing about how DeepSeek handles long projects, other languages or sensitive topics. Prices change, so check DeepSeek's page before you build on it.