DeepSeek API: Models, prices in CNY, free chat app and how to get a key

Plans, prices and features checked on the official pages on October 26, 2026.

The DeepSeek API gives developers two models, deepseek-flash and deepseek-v4-pro, through a format compatible with both the OpenAI and Anthropic SDKs. A key comes from the DeepSeek platform, and existing code only needs a new base URL. DeepSeek also offers a free chat on the web and in mobile and desktop apps.

Both API models list a 1M-token context window, and prices drop to half outside peak hours. The figures below come from DeepSeek’s own documentation, read on October 26, 2026.

Two API models cover everything DeepSeek sells today

DeepSeek’s docs list two model names. deepseek-flash is served by DeepSeek-V4.1-Flash-1017, released October 17, 2026. deepseek-v4-pro is served by DeepSeek-V4-Pro-0813.

The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the docs say those models are retired. Requests go to V4.1 Flash and are billed at the Flash price.

The pricing page lists these shared specs:

• Context length of 1M tokens and a maximum output of 384K.
• Thinking and non-thinking modes, with thinking as the default.
• JSON output, tool calls, the Responses API and the Anthropic API.
• Chat prefix completion and FIM completion, both in beta. FIM works only in non-thinking mode.
• Image understanding on Flash only. The page marks Pro as not supporting it.

The older names that show up in searches, such as R1, V3 and V3.2, appear on DeepSeek’s site as research releases. The API docs do not list them as callable models.

Prices halve off-peak, and the pricing page lists them in yuan

The pricing page lists prices in Chinese yuan (CNY) per million tokens. Peak hours are 9:00-12:00 and 14:00-18:00 Beijing time, Monday to Friday, excluding Chinese public holidays. Every other time, including weekends, is off-peak at half the price.

Prices checked on October 26, 2026. DeepSeek notes that prices may change. Charges come from your top-up balance or gift balance, and the gift balance is used first.

The Search and Fetch API is priced separately: $0.75 per 1,000 successful requests per endpoint, according to the guide (the pricing page lists 5 CNY). Only successful calls with usable results are billed. Calls made through DeepSeek Harness, Claude Code or Codex are free until the end of October 31, 2026, Beijing time, and billing starts after that.

Getting an API key and sending a first request takes five steps

1. Sign in at the DeepSeek platform and open the API keys page (platform.deepseek.com/api_keys).
2. Create a key and store it in an environment variable such as DEEPSEEK_API_KEY.
3. Point your SDK at https://api.deepseek.com for the OpenAI format, or https://api.deepseek.com/anthropic for the Anthropic format.
4. Set the model to deepseek-flash or deepseek-v4-pro.
5. Optionally set stream to true, reasoning_effort (the docs example uses "high"), or pass {"thinking": {"type": "enabled"}} to control thinking.

The docs say tools such as Claude Code, Codex and OpenCode can use DeepSeek as a backend without code changes. DeepSeek Harness, an agent harness, is in developer preview. The V4.1 Flash release notes recommend Harness v0.2.1-rc.1 or later, and say the model is also tuned for Tencent CodeBuddy and WorkBuddy.

Rate limits are about concurrency, not requests per minute

Each account has a cap on simultaneous requests per model: 2500 for Flash and 500 for Pro. The limit counts the whole account, whatever key is used. A request holds a slot from sending until the answer completes.

Go over the cap and the API returns an HTTP 429 error. DeepSeek says higher concurrency can be requested through a form, at no extra cost, matched to business need. While a request waits, non-streaming calls receive empty lines and streaming calls receive : keep-alive comments. If inference has not started after 10 minutes, the server closes the connection.

Search is limited to 100 requests per second per account and Fetch to 25, plus a per-domain limit on Fetch.

The free chat and apps run the latest flagship, but the fine print is thin

DeepSeek’s download page describes its app as a free way to chat with its models and as the official DeepSeek assistant. It lists these options:

• iPhone, through an App Store listing.
• Android, as a direct APK from DeepSeek’s download server rather than a Google Play link.
• Windows (x64) and macOS (Apple chip, macOS 13 or later) desktop apps, branded DeepSeek Harness.

The desktop version can run in the background and read and write local files, and DeepSeek says it is still being updated. The web chat sits at chat.deepseek.com. That page could not be read for this guide, so daily limits, if any, are not confirmed here.

The official pages do not describe image generation, image editing or video generation. What they do list is image understanding: V4.1 Flash is tagged image-text-to-text on Hugging Face and the pricing page shows image understanding as supported for Flash. Search results for “DeepSeek OCR”, “coder” and “R2” lead to separate releases that these pages do not cover, so they are left out here.

Open weights mean V4.1 Flash can be run locally, at a size

DeepSeek published the V4.1 Flash weights on Hugging Face and ModelScope under the MIT license. The model page lists 763B total parameters, a 1M-token context window and text plus image input. It does not give an active parameter count.

The page points to its inference folder and lists vLLM, SGLang and Docker Model Runner, with quantized builds for llama.cpp, Ollama and LM Studio. At 763B parameters, even quantized builds need serious hardware, so most individuals will use the API or the app. For how smaller models are trained from bigger ones, see our explainer on model distillation.

Data goes to servers in China, which matters for what you paste in

DeepSeek’s privacy policy names Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the operator. It says the company collects, processes and stores personal data in the People’s Republic of China.

The policy lists what it collects: account details, prompts, uploaded files, photos and chat history, device and network data, usage logs and approximate location from the IP address. Data is kept for as long as the account exists, and longer where needed for legal or business reasons. Deleting an account is permanent.

Whether that is acceptable depends on your own rules. Some governments have restricted AI chat tools on staff computers, as in Sweden’s Government Offices. Keeping confidential data out of any hosted chat, or running the open weights on your own hardware, avoids the question.

Alternatives worth a look if DeepSeek does not fit

Other open-weight and hosted models compete on price and coding skill. Our weekly roundups track them, including the week DeepSeek V4.1 Flash and Kimi K3.1 Pro led the releases. Check each maker’s own pricing page, because rates and peak rules differ.

Common questions