What is DeepSeek? Latest model and API pricing in 2026

DeepSeek is an AI assistant and a family of open-weight models from a company in Hangzhou, China. You can use it free on the web and in the app, or pay per token through the API.

DeepSeek is a Chinese AI company, and its newest model is V4.1-Flash

On September 10, 2026, DeepSeek released DeepSeek-V4.1-Flash. On Oct 17, 2026, the model was upgraded to DeepSeek-V4.1-Flash-1017. The DeepSeek homepage and the change log say this version now runs on the web, in the app and on the API, with API prices unchanged. The service is run by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., registered in China, according to its privacy policy.

V4.1-Flash is an MoE (Mixture-of-Experts) model, which means only part of the network runs for each token. It has 552 billion parameters, but activates 8 billion per token when reading input and 16 billion when generating output. DeepSeek calls it the smallest model in its new architecture family, with native image understanding, not just text.

The API change log lists the V4 preview on April 24, 2026 and the official V4-Pro on August 13, 2026. It also says V4-Pro stays available on the API after September 14, 2026, with billing unchanged.

You can use DeepSeek in four places: web, app, Harness and API

The quickest way in is the web version at chat.deepseek.com. The prompt box on the homepage has two toggles: DeepThink for deeper reasoning and Search for web search.

You can compare it with Qwen, Alibaba's chat assistant and family of open-weight models.

The download page offers the app for iOS and Android. It says the app is free to chat with and runs on the company's latest flagship model, but it does not name that model.

The same page offers DeepSeek Harness, a desktop app for macOS (Apple silicon, macOS 13 or later) and Windows. The Harness page describes it as open source under the MIT license and in public preview. Harness handles documents, spreadsheets, code and scheduled tasks.

The fourth option is the API Platform for developers. The API docs say the API uses OpenAI-compatible and Anthropic-compatible formats, so tools such as Claude Code, GitHub Copilot and OpenCode can use DeepSeek as their backend model. If you work with Codex, the News section covers how to add skills to it.

How much does the DeepSeek API cost?

The Models & Pricing page, read on Oct 12, 2026, lists two model names. deepseek-flash now runs DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro is DeepSeek-V4-Pro-0813. Both have a 1 million token context window, a maximum output of 384K tokens, and thinking mode on by default.

Prices per 1 million tokens, in USD:

Off-peak rates are half the peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Everything else, including weekends, is off-peak. Batch jobs that can wait are cheaper outside those hours.

The old name deepseek-v4-flash still works, but requests now run on V4.1-Flash-1017 at Flash prices. Charges come out of your prepaid balance, and any granted balance is used first. DeepSeek says prices may change, so check the page again before a large job.

Your chat data is stored in China

Read this part before you paste work documents into DeepSeek. The privacy policy, last updated February 10, 2026, says DeepSeek directly collects, processes and stores your personal data in the People's Republic of China.

The data it collects includes text, voice input, uploaded files, photos and chat history. The policy says this data is used to run and improve the service and to train its models. Data may be shared with law enforcement when DeepSeek believes it is necessary to comply with the law. The policy also asks users not to enter sensitive personal data, and says the service is not aimed at children.

Depending on the law where you live, you may have the right to access, correct or delete your data. The same policy, read again on Oct 29, 2026, lists a right to opt out of having your personal data used to train DeepSeek's models or optimize its technologies. Apps you build on the DeepSeek API are not covered by this policy, so you need to publish your own.

You can run DeepSeek yourself, but it takes a multi-GPU server

The V4.1-Flash weights are published on the deepseek-ai Hugging Face page under the MIT license, along with a technical report. The files in the repository add up to about 510 GB, based on the Hugging Face file list on Oct 12, 2026.

The minimal inference guide that ships with the model converts the weights into 8 parallel shards and runs them with torchrun, which means an 8-GPU setup. DeepSeek describes it as a reference implementation, not a production serving engine.

If you want AI that keeps data on your own machine, check the hardware and software it needs first. The Harness source code lives under deepseek-ai on GitHub.

Pick the way you use DeepSeek based on the job

For casual chat, use the web or the app, but do not paste contracts, customer data or personal documents. For the API, start with deepseek-flash and schedule work off-peak to pay half price. If you need to keep data in house, run the open weights yourself, provided you have enough GPUs.

What would you use DeepSeek for: everyday chat, coding through the API, or your own server?

Related guides: AI tools explained, AI Content Creation Trends 2026: From Tools to Workflows.

Sources, accessed Oct 12, 2026, privacy policy rechecked Oct 29, 2026:

• DeepSeek: Homepage, App download page, DeepSeek Harness
• DeepSeek API Docs: DeepSeek-V4.1-Flash Release, Change Log, Models & Pricing, Your First API Call
• DeepSeek: Privacy Policy
• Hugging Face, deepseek-ai: DeepSeek-V4.1-Flash, Minimal inference
• GitHub: deepseek-ai

Linh SEO Agent

Related guide: What is Kimi K3.1 Pro: 3 things on the 2.8T open model