What is DeepSeek? Is it free, is it safe, does it work in Turkish, and how much does the API cost?

DeepSeek is a product not of a software company but of an AI lab: Hangzhou DeepSeek Artificial Intelligence Co., Ltd. It answers the questions you type into the chat screen; it writes text and code, extracts text from files you upload, searches the web and thinks step by step before answering a hard question (DeepThink).

The same company runs three things that are easy to confuse. The chat app (web, iPhone, Android) is free. The API (for connecting to software) is pay-as-you-go. The open-weight models and DeepSeek Harness are options you download and use, with free software. For most people the chat app alone is enough.

The official terms list the output types as text, tables and code; generating photos or drawings is not on that list. Users also report that it doesn't generate images.

Who it helps: students, content creators, developers and small business owners who want a fast, simple, free chat. Who it doesn't suit: anyone who will work with customer, patient or employee data: the data is processed in China (see the Data section).

Chat needs no installation: you can use it in the browser. The mobile app works with the same account and chat history syncs across devices.

1. 1Open chat.deepseek.com or the app and sign in.
2. 2Type what you want. Attach a document or photo if you have one; the help page says there are limits on file format, number and size but gives no figures.
3. 3For a hard calculation, logic or planning question, turn on DeepThink; if you need current information, turn on Search (separate section below).
4. 4Read the answer and ask a follow-up. If you edit an earlier message in the same chat, the alternative answers are saved as "branches"; you switch between them with the arrows and numbers.

!DeepSeek can make mistakes. The official terms say the output may contain errors and should not be treated as medical, legal or financial advice. Check figures, dates and quotes against their source. Reported by users One user noticed that the rulings of Turkey's Court of Cassation cited in a legal article were not real; always verify case numbers.

The DeepSeek chat box has two buttons: DeepThink and Search. Both are free. The official announcement says "web search and Deep-Think mode"; knowing which one a question needs makes a noticeable difference to the result.

• When to use DeepThink? For work that needs step-by-step thinking, such as hard calculations, logic puzzles, planning and debugging code. According to the API documentation, in thinking mode the model produces a chain of reasoning before giving the final answer. The answer may take longer to start.
• When to use Search? For topics that change after the model's knowledge cutoff, such as current information, news, prices and regulations. Search uses a third-party search API, and the keywords in your input are sent to that service (Privacy Policy). With Search on, accuracy may improve somewhat, but false information can't be ruled out entirely (Terms of Use 5.2).

If you are going to search for something personal (a customer name, phone number, address), don't turn on Search: the keywords go to an outside search service.

Button names are from the official home page ("DeepThink", "Search"). How thinking mode works is described in DeepSeek's API documentation; the detailed behavior on the chat screen is not stated in official sources.

There are four separate places under the DeepSeek name. Which one you use depends on the job; for most people the first is enough.

They don't replace each other: chat runs on DeepSeek's servers and the data goes to China; if you run an open-weight model on your own hardware, the data stays on your machine (but that means a much bigger setup and more hardware).

DeepSeek Harness is DeepSeek's open-source agent app, currently in preview. It isn't question-and-answer like the chat app; it works in a folder on your computer, reads and edits files, runs commands and carries out long jobs in the background. Everything is designed as a "plugin"; you add the plugins you want or have it write a new plugin through chat.

• Where: desktop app for macOS (Apple silicon, macOS 13 or later) and Windows (64-bit); also a web interface via Node.js with the npx @deepseek-ai/dsh web command.
• License and cost: the software is open source (MIT) and free. To use the DeepSeek model you enter a DeepSeek API key under Settings› then Models; usage is billed at API prices. Another provider or an OpenAI-compatible endpoint can also be connected. The official page doesn't say whether Harness includes any free model allowance of its own.
• Status: preview; core plugins and APIs are changing. Back up your files before using it for production work.

!Harness is an agent: it can write to your files and run commands. Don't open important folders directly; work in a test folder first and read the approval requests.

The DeepSeek API speaks the OpenAI and Anthropic formats. That means it can be plugged into coding agent tools as the backend model: you switch the model to DeepSeek without changing the tool itself. DeepSeek has published official guides for these tools:

Claude Code as an example (commands from the official guide, macOS and Linux; on Windows the same variables are written with $env:):

1. 1After setting these variables, go to your project folder and type claude.
2. 2If Claude Code isn't installed, the official guide gives the command npm install -g @anthropic-ai/claude-code with Node.js 18+; on Windows you also need Git for Windows.

• Cost: Usage is billed at DeepSeek API prices and is not included in a Claude subscription. If web search is used in Claude Code, the extra model request for summarizing is also billed (official guide).
• Model mapping: Claude Code or Claude Desktop names are translated to DeepSeek names: "claude-opus" names go to deepseek-v4-pro (at V4 Pro prices), "claude-haiku" and "claude-sonnet" names go to deepseek-flash.

!Your API key is like a password: don't write it into code, a browser or a file you share. If it leaks, revoke the key immediately and create a new one, as the DeepSeek help page describes.

Open its card

You can download the weights of DeepSeek models from Hugging Face and run them on your own hardware; DeepSeek releases them under the MIT license. The software is free, but the server/hardware cost and updates are on you.

• The newest model (V4.1-Flash) has 552 billion parameters. For large-scale deployment, DeepSeek gives the example of "a cluster of 2,000 GPUs and storage"; it won't fit on an ordinary computer. The official model card doesn't list hardware requirements for an ordinary computer.
• Small "distilled" models: the R1 model card lists six small models derived from DeepSeek-R1, ranging from 1.5 billion to 70 billion parameters. These are separate models based on Qwen and Llama; they are not the same as the DeepSeek in the app. Reported by users Users report that answers get weaker with a small local model and that Chinese characters sometimes slip in.
• Data: the Privacy Policy applies only to DeepSeek's own apps and services; it excludes apps written by other developers. If you run the model on your own computer, the data doesn't go to DeepSeek's servers; if you run it through another provider, that provider's rules apply.

Setting up the programs used to run models on your own computer is covered in separate guides:

Open its card

Open its card

Sources: Hugging Face model cards (V4.1-Flash, R1), DeepSeek V4.1-Flash announcement (September 10, 2026). Hardware requirements are not in the official sources; check which model fits your computer in the relevant program's guide and on the model page.

DeepSeek has several models and their names change often. In the chat app you usually don't pick the model yourself; model choice mainly matters for the API, Harness and open-weight use. The table below is based on DeepSeek's own descriptions.

Two official pages describe the future of the V4 Pro API differently: the September 10 news page says V4 Pro requests will be redirected to V4.1 Flash from September 14; the same site's changelog says the V4 Pro API service will continue, and the pricing page still lists V4 Pro. This section was checked on October 14, 2026 (sources: DeepSeek API pricing page, changelog, announcements of April 24, August 13 and September 10, 2026, Hugging Face model cards); model versions change often.

Chat is free: the web and mobile apps, including DeepThink, Search and file uploads. The official announcement says "no ads, no in-app purchases"; the App Store Turkey page also shows it as free and lists no in-app purchases for DeepSeek. There is no monthly subscription plan on the official pricing page. The only paid thing is the API.

Prices were checked on October 14, 2026 (api-docs.deepseek.com › Models & Pricing). DeepSeek reserves the right to change prices.

Official pricing page last checked on October 28, 2026.

If the answer comes back in the wrong language, add "answer in Turkish" to your message or regenerate the answer.

!Don't paste text that names customers, students, patients or employees into DeepSeek. If you need to work with personal data, mask the data or run the model on your own hardware.

For Turkey, no official decision or KVKK announcement specific to DeepSeek appears, and the app is live on App Store Turkey. This is not a legal guarantee; the status is based on official sources as of October 14, 2026. In other countries, data protection authorities took these steps in 2025:

Sources: the authorities' own announcements (Garante decision 10098477, PIPC press release, Berlin BlnBDI press release and 2025 annual report). Their current status (e.g. the app being republished in South Korea) is not clear in official statements. News sites were not counted as sources in this table.

If you work at a public institution or a company, check your organization's AI usage policy: even without a legal ban, company policy may restrict DeepSeek.

According to the terms of use, you keep the rights to your inputs; the rights to the outputs (if any) are assigned to you. You can use the output for a wide range of purposes, including personal use, academic research, developing derivative products and even training another model, as long as it is legal and you follow the terms.

• If you publish the output, you need to check its accuracy and state that it is AI-generated (Terms 3.1).
• For significant decisions about people, such as credit, education, hiring, insurance, legal and health matters, the output must go through human review; the responsibility is yours (Terms 5.4).
• The open-weight model repositories are MIT-licensed (V4.1-Flash, V4-Pro, V4-Flash, V3.2, R1): you are free to use them, modify them and use them in a commercial product; you must keep the license text and copyright notice. Exception: R1's small distilled models derive from bases under other licenses: the Qwen-based ones are Apache 2.0, while Llama-8B and Llama-70B are bound by Llama's own licenses.

Source: DeepSeek Terms of Use (March 27, 2026) 3.1, 3.6, 4.2, 5.4; Hugging Face model cards. This is not legal advice.

Since chat is free for everyone, there is no special student plan or discount. The Terms of Use say the service is mainly intended for adults: people under 18 (or under the minimum age in their country) should read the terms with a guardian and use it with the guardian's permission. The App Store Turkey page shows an age rating of 13+.

If you want to work with it as an organization, you can request an offline agreement for the API with a "Cooperation Agreement Application" ticket; a "Rate Limit Increase Request" for higher concurrent requests is free.

There is no monthly subscription, so there is no plan to cancel. The only things to delete are the account and the chats.

1. 1Back up your data first: Settings› then Data› then Export data. The download link is valid for 7 days.
2. 2Open Settings› then Account settings› then Delete Account (if it doesn't appear in an old version, update the app).
3. 3Confirm. The account's entire chat history is permanently deleted, any API balance is lost and the account can't be reopened.

DeepSeek states in Terms of Use 2.5 that it may keep some data after deletion for legal obligations.

The list below shows each feature separately: whether it is built into the app, needs a connection or a separate installation, and whether it is free. A free product doesn't mean every feature is free.

DeepSeekFree, simple chat with DeepThink and Search. Data is processed in China.ChatGPTImage generation and voice chat in one place.ClaudeStrong with long documents, files and code; doesn't generate photos or video.GeminiImages and Deep Research on the free plan; works with Google apps.

The rows for the other tools are summaries from those tools' own guides. Reported by users On Turkish forums, DeepSeek's speed and simplicity are praised, while its lack of image generation and its refusal to answer some political questions are criticized.