What is DeepSeek?

Who makes DeepSeek, what the chat and reasoner models are good at, why the open weights matter for your bill, and the honest limits.

In this lesson you will

Explain what DeepSeek is and what makes it different from ChatGPT, Claude and Gemini
Describe the chat model vs reasoner model split in one sentence
Know why open weights change the cost conversation
List the situations where DeepSeek is not the right pick

DeepSeek is a family of large language models built by DeepSeek, a Chinese AI lab. It is best known for two things: model quality that competes with the big US labs on many tasks, and API prices that are a fraction of what those labs charge. That second point is why this course exists on a site about not overpaying for AI.

You can use DeepSeek three ways. There is a free web and mobile chat app at chat.deepseek.com, a pay-as-you-go API at platform.deepseek.com, and — because DeepSeek publishes its model weights under permissive licences — you can run the models through other hosting providers or on your own hardware. Most other frontier labs offer the first two only.

Figure 1.Every DeepSeek model works the same way underneath: text is split into tokens, the model predicts the next token, and you pay per token in and per token out on the API.

Chat models and reasoner models

DeepSeek ships two kinds of model. The chat line (DeepSeek V3, V3.1, V3.2 and the V4 family) answers directly, the way you expect a chatbot to. The reasoner line (DeepSeek R1 and its successors) thinks first: it generates a long hidden chain of reasoning before it writes the answer, which helps on maths, logic and multi-step coding but takes longer and produces far more output tokens.

In the app this shows up as a DeepThink toggle. On the API it is the difference between the deepseek-chat and deepseek-reasoner model names. The cost lesson is simple: reasoning tokens are output tokens, and output tokens are the expensive ones. Lesson 2 covers when the extra thinking is worth paying for.

Why open weights matter for your bill

Figure 2.Open weights mean the same model can be served by DeepSeek, by third-party hosts, or by you. Price competition between hosts is what keeps DeepSeek cheap.

DeepSeek releases its model weights under MIT-style licences. That means any cloud provider can download the weights and sell access, and you can too. When several hosts serve the same model, they compete on price, so the going rate for a DeepSeek model tends to fall over time. It also means there is no lock-in: if DeepSeek's own API is slow or unavailable, the identical model is one base-URL change away on another host.

What DeepSeek is good at

Coding. The V3 and V4 chat models are strong at writing and editing code, and the reasoner is good at debugging and algorithmic problems. See DeepSeek vs Claude for coding.
Maths and structured reasoning. R1-style models were built for this and it is where the DeepThink mode earns its extra tokens.
High-volume text work. Summarising, classifying, extracting and translating large amounts of text is where the low API price compounds into real savings.
Chinese and English. DeepSeek is trained heavily on both; quality in other languages is decent but less consistent.

Honest limits

DeepSeek is not the answer to everything. The official API has had capacity problems during demand spikes, so latency and availability are less predictable than the incumbents. The chat app has fewer polished extras than ChatGPT or Gemini: no equivalent of a mature projects, memory or connected-apps ecosystem. Image understanding is limited to specific vision variants, and there is no image generation.

The rest of this course is organised around one question: for the task in front of you, which of the three ways to use DeepSeek is cheapest without being worse? For most beginners the answer starts with the free app, and Lesson 3 shows exactly where the free app stops being enough.

Knowledge check

Which statement about DeepSeek's reasoner models (R1 style) is correct?

Lesson FAQ

Is DeepSeek really free?

The web and mobile chat app is free with no paid consumer tier recorded in our catalogue. The API is pay-as-you-go. Running the open weights yourself is free in licence terms but you pay for the hardware or cloud compute.

Is DeepSeek open source?

The model weights are released under MIT-style licences, which is why third-party hosts and self-hosting exist. Training data and full training code are not released, so the accurate term is open-weight rather than fully open source.

How does DeepSeek compare with ChatGPT and Claude?

On price, DeepSeek's API is far cheaper. On quality it is competitive for coding, maths and text work, and behind on ecosystem features and multimodal input. Our [three-way comparison](/compare/deepseek-vs-chatgpt-vs-claude) goes through it task by task.