DeepSeek is a Chinese AI lab that offers a free chat app and a very low-cost API, both running its own open-weight models. As of September 2026 its line-up is DeepSeek-V4.1-Flash and DeepSeek-V4-Pro, each with a 1-million-token context window and MIT-licensed weights. It is strong value for developers, but data stored in China and government restrictions rule it out for sensitive work.
Last updated: September 2026
What is DeepSeek?
DeepSeek is developed by Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., founded in July 2023 by Liang Wenfeng and backed by the Chinese hedge fund High-Flyer. Its products are DeepSeek Web (chat.deepseek.com), the DeepSeek mobile app, the developer API platform and DeepSeek Harness, an open-source agent framework.
Two models are current as of September 2026, according to DeepSeek’s API pricing page:
- DeepSeek-V4.1-Flash (released 10 September 2026): a mixture-of-experts model with 552 billion backbone parameters that activates about 8 billion parameters per token for input and 16 billion for output. It adds native image understanding, and DeepSeek says tests by multiple parties put it ahead of V4-Pro on performance, cost and speed.
- DeepSeek-V4-Pro: 1.6 trillion total parameters with 49 billion active, released as a preview in April 2026 and made generally available on 13 August 2026.
DeepSeek first announced that V4-Pro API requests would be routed to V4.1-Flash from 14 September 2026 until a V4.1-Pro launches. Its changelog then stated it would keep serving V4-Pro after that date with billing unchanged. V4.1-Pro is not yet listed on the pricing page.
From R1 to V4.1: what changed since January 2025
| Date | Release or event | Why it matters |
|---|---|---|
| 20 Jan 2025 | DeepSeek-R1 | Open reasoning model that DeepSeek said performed comparably to OpenAI o1, under the MIT license |
| 26–27 Jan 2025 | App hits No. 1 on the US App Store | Nvidia lost $589 billion in market value on 27 January |
| 28 May 2025 | R1-0528 | Stronger reasoning and function calling |
| 21 Aug 2025 | DeepSeek-V3.1 | One hybrid model with thinking and non-thinking modes, toggled by the app’s DeepThink button |
| 1 Dec 2025 | DeepSeek-V3.2 | Upgrade to both API models, following the V3.2-Exp test in September |
| 24 Apr 2026 | DeepSeek-V4 preview | V4-Pro and V4-Flash, 1M-token context, open weights |
| 13 Aug 2026 | V4-Pro general availability | Better agent performance, low/high/max thinking effort, peak and off-peak pricing |
| 10 Sep 2026 | DeepSeek-V4.1-Flash | New architecture, native vision, lower API prices |
Key features of DeepSeek
Free chat app with Instant and Expert modes
The web chat and iOS and Android apps are free. Since April 2026, the chat has offered two modes: Instant for fast everyday replies and Expert for harder problems. A web search toggle pulls in live results. At the modes’ April 2026 rollout, Cocoloop reported that file uploads worked only in Instant mode, so check which mode you are in before attaching a PDF.
1-million-token context window
Both API models accept up to 1 million tokens of context and can return up to 384,000 output tokens. In practice, this means you can pass a whole codebase, a long contract set or a year of reports in a single request.
Thinking and non-thinking modes
Since V3.1, one DeepSeek model handles both quick answers and step-by-step reasoning. On the API, thinking is the default and can be turned off. The August 2026 V4-Pro update added low, high and max effort levels.
Native image understanding
V4.1-Flash reads images as well as text. Through the API, deepseek-flash accepts JPEG, PNG, GIF and WebP images, up to 600 per request, to describe pictures, read screenshots or analyze charts. V4-Pro does not support vision.
OpenAI- and Anthropic-compatible API
The API uses OpenAI-compatible and Anthropic-compatible formats and supports JSON output, tool calls and the OpenAI Responses API. DeepSeek’s docs say you can plug it into coding agents such as Claude Code, GitHub Copilot and OpenCode as the backend model without writing code.
Open weights under the MIT license
DeepSeek publishes its model weights on Hugging Face. R1, V4-Pro and V4.1-Flash all use the MIT license, which permits commercial use and derivative models. Companies can self-host them instead of sending data to DeepSeek.
Context caching and off-peak pricing
Repeated input that hits DeepSeek’s cache is billed at a small fraction of the normal input rate. Since August 2026, off-peak rates are half the peak rates, which rewards batch jobs scheduled outside weekday peak hours.
DeepSeek Harness (developer preview)
Released in August 2026 under the MIT license, DeepSeek Harness is an open-source runtime for building AI agents from plugins. DeepSeek warns it is a developer preview with compatibility-breaking changes ahead.
Benefits of using DeepSeek
- Very low API cost: V4.1-Flash output costs $0.60 per million tokens off-peak.
- Free consumer app: no subscription needed for the web or mobile chat.
- Self-hosting option: MIT-licensed weights let you run the models on your own infrastructure.
- Easy migration: OpenAI- and Anthropic-compatible endpoints mean most existing code and agent tools work after a base-URL change.
- Long context: 1 million tokens covers large documents and repositories in one pass.
DeepSeek use cases by role
- Developers: a low-cost backend for coding agents, code review and long-context repository analysis.
- Startup founders and SaaS builders: high-volume features such as summarization, classification and chat support where token costs drive margins.
- ML engineers and researchers: studying and fine-tuning open-weight frontier models under a permissive license.
- Students: free help with maths, coding and explanations in the app, keeping personal data out of prompts.
- Enterprises: self-hosted DeepSeek weights for internal workloads, avoiding the hosted service’s data transfer to China.
How much does DeepSeek cost?
The DeepSeek chat app is free. The API is pay-as-you-go. The table shows API prices per million tokens as of September 2026, from DeepSeek’s pricing page, in USD.
| Model (API name) | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| DeepSeek-V4.1-Flash (deepseek-flash), off-peak | $0.003 | $0.15 | $0.60 |
| DeepSeek-V4.1-Flash (deepseek-flash), peak | $0.006 | $0.30 | $1.20 |
| DeepSeek-V4-Pro (deepseek-v4-pro), off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek-V4-Pro (deepseek-v4-pro), peak | $0.044 | $1.32 | $3.96 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is billed at the off-peak rate. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to V4.1-Flash at Flash prices. The legacy deepseek-chat and deepseek-reasoner names were set to be fully retired after 24 July 2026 and no longer appear on the pricing page, so update older integrations.
DeepSeek pros and cons
Pros
- Very low per-token API prices, with cache and off-peak discounts
- Free web and mobile chat
- MIT-licensed open weights for self-hosting and fine-tuning
- 1-million-token context and up to 384,000 output tokens
- Drop-in compatibility with OpenAI and Anthropic SDKs and popular coding agents
Cons
- Hosted-service data is stored in China
- Banned on government devices in several countries and regions
- The hosted app has deflected politically sensitive Chinese topics in press tests
- Fast-changing model names and routing can break integrations
- Regulators in several countries have acted over its data transfers
Privacy, censorship and government restrictions
DeepSeek’s privacy policy (updated 10 February 2026) says it collects, processes and stores personal data in the People’s Republic of China and may use it to train its models, with a right to opt out. In January 2025, Wiz researchers found an exposed DeepSeek database with over a million log lines, including chat history and API keys. DeepSeek secured it promptly after disclosure.
In tests reported by CBC News, the app declined questions on Tiananmen Square and Taiwan’s status with “Sorry, that’s beyond my current scope.” Separately, Anthropic alleged in February 2026 that DeepSeek had generated over 150,000 exchanges with Claude to extract its capabilities. The main government actions are:
| Country or body | Action | Date |
|---|---|---|
| Italy | Privacy regulator (Garante) ordered the service blocked; a separate consumer probe closed in January 2026 after DeepSeek committed to clearer hallucination warnings | 30 Jan 2025 |
| Taiwan | Government agencies and critical infrastructure barred from using DeepSeek | Early Feb 2025 |
| Australia | Banned on federal government devices | 4 Feb 2025 |
| New York State | Banned on state government devices and networks | 10 Feb 2025 |
| South Korea | New downloads suspended; regulator later found user data had been transferred abroad without consent; downloads resumed on 28 April | 17 Feb 2025 |
| Germany | Berlin’s data protection commissioner asked Apple and Google to remove the app under the Digital Services Act | Jun 2025 |
| Czech Republic | All DeepSeek products banned in state administration | Jul 2025 |
| United States | FY2026 NDAA requires removal of DeepSeek AI from Department of Defense systems, with waivers for research | 18 Dec 2025 |
How to use DeepSeek
- Open chat.deepseek.com or install the DeepSeek app from the App Store or Google Play, then create an account.
- Pick Instant mode for quick answers or Expert mode for complex reasoning, and switch on search if you need current information.
- Keep personal, client and confidential data out of your prompts, and use the training opt-out described in DeepSeek’s privacy policy if you prefer.
- For the API, sign up on the DeepSeek API platform, top up your balance and create an API key.
- Point your OpenAI SDK at https://api.deepseek.com (or your Anthropic SDK at https://api.deepseek.com/anthropic) and set the model to deepseek-flash or deepseek-v4-pro.
- Reuse fixed prompt prefixes to benefit from cache-hit pricing, and schedule batch jobs outside peak hours.
- For sensitive workloads, download the weights from Hugging Face and self-host instead of using the hosted service.
DeepSeek alternatives
- Qwen: pick Alibaba’s family if you want a mix of Apache 2.0 open-weight models and a free chat app from another major Chinese lab.
- Kimi: pick Moonshot AI’s assistant, now on Kimi K3, for agent swarms, deep research and document or slide generation.
- ChatGPT: pick it if you want a generous free tier plus Business and Enterprise plans with workspace admin settings and SSO.
- Claude: pick it for careful writing and coding, with Claude Code on paid plans and Team and Enterprise plans with SSO and admin controls.
- Mistral Vibe: Mistral AI’s assistant, formerly Le Chat. Pick it if you need on-premises or private-cloud deployment with full data residency.
Frequently asked questions
Is DeepSeek free?
Yes, for the consumer product. The DeepSeek web chat and the iOS and Android apps are free, and the App Store listing shows no paid tier. Developers pay separately for the API, billed per million tokens. As of September 2026, DeepSeek-V4.1-Flash costs $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, with rates doubling during weekday peak hours.
What is the latest DeepSeek model?
As of September 2026, the newest model is DeepSeek-V4.1-Flash, released on 10 September 2026 with native image understanding and a 1-million-token context window. DeepSeek-V4-Pro, which reached general availability in August 2026, is still offered on the API. DeepSeek has referred to a coming V4.1-Pro, but it is not yet listed on the API pricing page.
What happened to DeepSeek-R1?
DeepSeek-R1, released on 20 January 2025, was the reasoning model behind DeepSeek’s breakthrough. It was updated in May 2025, then folded into DeepSeek-V3.1 in August 2025, which combined thinking and non-thinking modes in one model. Its API name, deepseek-reasoner, was set to be fully retired after 24 July 2026. R1’s MIT-licensed weights remain on Hugging Face.
Is DeepSeek open source?
DeepSeek publishes open weights. DeepSeek-R1, DeepSeek-V4-Pro and DeepSeek-V4.1-Flash are all on Hugging Face under the MIT license, which allows commercial use and modification. Open weights means the trained model files are public; the hosted chat app and API remain proprietary services run by DeepSeek.
Is DeepSeek safe to use?
It depends on what you share. DeepSeek’s privacy policy says personal data is collected, processed and stored in China and may be used for model training, with an opt-out. Regulators in Italy, South Korea and Germany have acted over its data practices, and several governments ban it on official devices. Avoid entering confidential or personal data into the hosted app.
Is DeepSeek banned?
There is no general consumer ban in most countries, but restrictions are widespread. Italy’s privacy regulator ordered it blocked in January 2025. Australia, Taiwan, New York State and the Czech Republic banned it on government devices or in public administration, and the US FY2026 NDAA requires its removal from Department of Defense systems. Check local rules and your employer’s policy.
Can I run DeepSeek locally?
Yes, because the weights are public, but the flagship models need server-class hardware: DeepSeek-V4-Pro has 1.6 trillion total parameters. Running the open weights on your own servers keeps prompts off DeepSeek’s infrastructure, which addresses the data-location concern. For laptops, the smaller R1 distilled models (1.5B to 70B parameters, based on Qwen and Llama) are the practical option.
Verdict
DeepSeek is one of the best-value AI options for developers in September 2026. V4.1-Flash delivers a 1-million-token context, vision and agent-ready APIs from $0.15 per million uncached input tokens off-peak, and the MIT-licensed weights let you self-host. The free app is capable for everyday use. But the hosted service stores data in China, and bans are widespread, so keep sensitive work on self-hosted weights or another provider.
Sources
- DeepSeek API Docs: Models & Pricing
- DeepSeek API Docs: Change Log
- DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient (DeepSeek API Docs)
- DeepSeek V4 Preview Release (DeepSeek API Docs)
- DeepSeek-V3.1 Release (DeepSeek API Docs)
- DeepSeek API Docs: Your First API Call
- DeepSeek API Docs: Vision
- DeepSeek official website
- DeepSeek Privacy Policy
- DeepSeek – AI Assistant (Apple App Store)
- deepseek-ai/DeepSeek-V4.1-Flash (Hugging Face)
- deepseek-ai/DeepSeek-V4-Pro (Hugging Face)
- deepseek-ai/DeepSeek-R1 (Hugging Face)
- deepseek-ai/deepseek-harness (GitHub)
- The open-sourcing of DeepSeek Harness (InfoQ)
- DeepSeek releases V4.1-Flash, says it outperforms flagship V4-Pro (SiliconANGLE)
- DeepSeek launches Fast and Expert chat modes (Cocoloop)
- Feature request: Replace Expert Mode with DeepThink toggle and fix search UX (DeepSeek GitHub issue #1381)
- DeepSeek (Wikipedia)
- DeepSeek displaces ChatGPT as the App Store's top app (TechCrunch)
- Nvidia stock plummets, loses record $589 billion as DeepSeek prompts questions over AI spending (Yahoo Finance)
- Wiz Research uncovers exposed DeepSeek database leaking sensitive information (Wiz)
- DeepSeek seems to struggle with questions that would upset Chinese authorities (CBC News)
- Italy bans Chinese DeepSeek AI over data privacy and ethical concerns (The Hacker News)
- Italian regulators end DeepSeek 'hallucination' probe (Silicon UK)
- South Korea removes DeepSeek from app stores pending privacy review (Al Jazeera)
- Gov't confirms DeepSeek's unauthorized data transfer abroad (The Korea Times)
- DeepSeek AI service reopens to South Korean users after suspension (Profit by Pakistan Today)
- DeepSeek dilemma: Taiwan's public sector ban (Radio Free Asia)
- DeepSeek banned from Australian government devices (The Register)
- Governor Hochul issues statewide ban on DeepSeek on government devices (New York State)
- Germany asks Google, Apple to remove DeepSeek AI from app stores (BleepingComputer)
- Czech Republic bans Chinese AI startup DeepSeek in government work (Euronews)
- FY2026 NDAA, Section 1532: Guidance and prohibition on use of certain artificial intelligence (ETO AGORA)
- Detecting and preventing distillation attacks (Anthropic)
- Qwen (Wikipedia)
- Qwen Chat
- Kimi (Moonshot AI)
- Le Chat is now Vibe (Mistral Help Center)
- Mistral Vibe (formerly Le Chat)
- ChatGPT pricing (OpenAI)
- SSO for ChatGPT Business: FAQ (OpenAI Help Center)
- Claude pricing (Anthropic)
