DeepSeek — an AI company with open models and a cheap API

• Who makes it

  An AI research company in Hangzhou, China, started in 2023 by the quantitative investment firm High-Flyer. Its founder is Liang Wenfeng.

• Open weights

  The model files themselves are published under the MIT license. You may download them, run them on your own server, change them and use them commercially.

• Model you call today

  DeepSeek-V4.1-Flash-1017 (the API name is deepseek-flash). It reads 1 million tokens of context and can also read pictures.

• API format

  The same format as the OpenAI and Anthropic APIs. Programs written for them work after you change only the address, key and model name.

Dates from DeepSeek's API change log. Each step made the model stronger or cheaper; R1 was the moment the whole world noticed.

1. 2024-05-17

   DeepSeek-V2 — an MoE model that already sold at a very low price

2. 2024-12-26

   DeepSeek-V3 — a large MoE model close to the best paid models

3. 2025-01-20

   DeepSeek-R1 — a reasoning model that "thinks before it answers". 671B parameters in total, 37B used per token, MIT licence. It matched top models in maths and coding at a fraction of the cost, and shook the AI industry

4. 2025-08-21

   DeepSeek-V3.1 — thinking and non-thinking modes in one model

5. 2025-12-01

   DeepSeek-V3.2 — cheaper long context

6. 2026-04-24

   DeepSeek-V4 — two lines: Pro (larger) and Flash (faster and cheaper)

7. 2026-09-10

   DeepSeek-V4.1-Flash — a stronger Flash

8. 2026-10-17

   DeepSeek-V4.1-Flash-1017 — better scores with shorter answers; tuned for agent work such as building whole projects

Prices are US dollars per 1 million tokens. During peak hours they double. A "cache hit" is the part of the input that is the same as the beginning of an earlier request.

Peak hours: 01:00–04:00 · 06:00–10:00 UTC, Monday to Friday. Weekends and Chinese public holidays are off-peak all day.

Create an API key on the DeepSeek platform, then send the same request you would send to OpenAI — only the address and model name change.

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
        "model": "deepseek-flash",
        "messages": [{"role": "user", "content": "Hello!"}],
        "thinking": {"type": "disabled"}
      }'

Thinking mode is on by default: the model first writes its reasoning, then the answer. For quick answers, such as a chatbot, turn it off with "thinking": {"type": "disabled"}.

Tools built for Claude (Anthropic format) use the address https://api.deepseek.com/anthropic instead.

• Automatic cache

  If the start of a request is the same as before (the system prompt, the earlier conversation), that part is a cache hit and costs $0.003 instead of $0.15 per million tokens — 50 times cheaper. You do not have to turn anything on.

• Off-peak hours

  Outside 01:00–04:00 · 06:00–10:00 UTC on weekdays, the price is half. Batch jobs that can wait are cheaper at night in Asia or at weekends.

• MoE design

  The model has many experts but computes each token with only a few of them, so it costs less to run than a dense model of the same size.

• Shorter answers

  The 1017 version finishes the same work with shorter output. You pay for output tokens, so a shorter answer is a cheaper answer.

• Empty answers in thinking mode

  In thinking mode, max_tokens counts the reasoning too. If it is too small, the reasoning uses it all and the answer comes back empty without any error. Raise max_tokens or turn thinking off.

• Old model names

  The old names deepseek-v4-flash, deepseek-v4-flash-vision-exp are still accepted, but those models are retired and deepseek-flash answers instead. Use deepseek-flash in new code.

• Where data goes

  DeepSeek's privacy policy says its servers are in the People's Republic of China. Do not send personal information or company secrets you are not allowed to send abroad.

• Prices change

  The prices on this page were checked on 2026-10-24. Check the official price list before you build a budget on them.

In one line: DeepSeek is "strong, cheap and open" — call it through the API for convenience, or download the weights when you need your own server.

This page is a plain-language summary of the official documents below, checked on 2026-10-24. If they differ, the official documents are right.

• DeepSeek API Docs — Your First API Call(opens in a new tab)
• DeepSeek API Docs — Models & Pricing(opens in a new tab)
• DeepSeek API Docs — Change Log(opens in a new tab)
• DeepSeek API Docs — DeepSeek-V4.1-Flash-1017 (2026-10-17)(opens in a new tab)
• DeepSeek API Docs — Thinking Mode(opens in a new tab)
• DeepSeek API Docs — Context Caching(opens in a new tab)
• Hugging Face — deepseek-ai/DeepSeek-V4.1-Flash-1017(opens in a new tab)
• Hugging Face — deepseek-ai/DeepSeek-R1(opens in a new tab)
• DeepSeek Privacy Policy(opens in a new tab)