DeepSeek
DeepSeek is a Chinese AI lab offering a free general-purpose chat assistant on web, iOS and Android, plus a low-cost API whose models ship with open MIT-licensed weights and a one-million-token context window.
What is DeepSeek?
DeepSeek is the public face of a Chinese research team that describes itself as dedicated to exploring AGI and committed to an open-source approach. It reaches users through three doors. The first is a free chat assistant at chat.deepseek.com, with companion applications for iOS and Android and a direct APK download; there is no paid consumer tier to upgrade to. The second is an open platform where developers create an account, generate a key and call the models directly. The third is the models themselves: DeepSeek publishes all weights, parameters and inference tool code under the permissive MIT License, together with a technical report for each release, so anyone can download and deploy them without asking permission.
Two models are served through the API. deepseek-v4-flash, currently at version DeepSeek-V4-Flash-0731, is the entry point; deepseek-v4-pro is the larger one. Both accept a context of one million tokens and return up to 384,000, both support a thinking mode that is on by default and can be switched off, and both expose JSON output, tool calls and context caching. The unusual part is the plumbing: the same key works against an OpenAI-format endpoint at api.deepseek.com and an Anthropic-format endpoint at api.deepseek.com/anthropic, so existing code moves over by changing a base URL rather than by being rewritten.
That compatibility is what the documentation builds on. DeepSeek publishes integration guides for Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix and CodeBuddy, and maps Claude model names onto its own — claude-opus becomes deepseek-v4-pro, claude-sonnet and claude-haiku become deepseek-v4-flash — so a coding agent can be redirected with two or three environment variables. Web search is supported natively on the API side.
The company behind it is Hangzhou DeepSeek Artificial Intelligence Co., Ltd. Its marketing site is deliberately spare: a home page, a transparency centre listing model cards and release dates, and nothing resembling an about, contact or features page. The substance sits in the legal documents and the API reference instead.
What it does
- Hold multi-turn conversations from text, image and file inputs, returning text, tables and code
- Reason step by step in a thinking mode that is enabled by default and tunable through a reasoning-effort setting
- Write and complete code, including fill-in-the-middle and chat prefix completion
- Call external tools and return strictly structured JSON
- Search the web, both inside the product and natively through the API
- Act as the backend model for coding agents such as Claude Code, Codex and OpenCode
- Work across a one-million-token context and cache it to cut the cost of repeated input
When to use DeepSeek / When not to
A quick filter to help you decide if DeepSeek is the right fit.
When to use DeepSeek
- Developers who want a cheap drop-in model endpoint, since the same key answers in both OpenAI and Anthropic request formats
- Users of terminal coding agents such as Claude Code, Codex, OpenCode or GitHub Copilot, which DeepSeek documents as supported backends
- Cost-sensitive engineering teams, where USD 0.14 per million input tokens changes what is worth automating
- Machine-learning researchers and self-hosting teams, who get weights, parameters and inference code under the MIT License
- Everyday users who want a capable assistant at no cost on the web or on an iOS or Android phone
When not to use DeepSeek
- Organisations bound to keep personal data inside the EU or the United States, since DeepSeek states it stores data in the People's Republic of China
- Procurement and compliance teams who require a signed data processing agreement, a named subprocessor list or a SOC 2 or ISO 27001 certificate, none of which the site publishes
- Anyone seeking professional medical, legal or financial advice, which the terms explicitly exclude from the value of any output
- Buyers who need an enterprise contract, a service-level agreement or a named account contact, as no sales channel is published at all
- Children and unsupervised minors, since the service is described as primarily intended for adults and requires a legal guardian's consent below eighteen
How to use DeepSeek
A typical end-to-end flow, from setup to results.
- Open chat.deepseek.com, or install the DeepSeek app from the App Store or Google Play
- Create an account with an email address, or sign in through a third-party service such as Apple or Google
- Type a prompt — plain text, an image or an uploaded file — and read the answer
- Turn on the Search function when the question needs current information
- Open the settings to review, copy or delete your chat history
- Switch off "Improve the model for everyone" in the settings if you do not want your inputs used for training
- For development work, register on platform.deepseek.com and generate a key under API keys
- Point your code at api.deepseek.com for the OpenAI format, or api.deepseek.com/anthropic for the Anthropic format, and pick deepseek-v4-flash or deepseek-v4-pro
- To redirect a coding agent, export ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, then launch it as usual
- Top up a balance before making paid calls; any granted balance is spent before the topped-up one
Pros & Cons
Pros
- The chat assistant is free on web and mobile, with no paid tier to upgrade to
- API pricing is aggressive: USD 0.14 per million input tokens and USD 0.28 per million output tokens on the entry model
- Cached input costs USD 0.0028 per million tokens, fifty times less than uncached input
- Weights, parameters and inference code are published under the MIT License, so self-hosting and distillation are permitted
- Dual OpenAI and Anthropic compatibility turns migration into a configuration change rather than a rewrite
- The terms assign output rights to the user and explicitly allow reuse, including training other models
- Documentation is unusually candid: model mechanism, training methods, a public training-data summary and model cards are all published
Cons
- Personal data is collected, processed and stored in the People's Republic of China, under mainland Chinese law and jurisdiction
- No data processing agreement, no named subprocessors and no security certification such as SOC 2 or ISO 27001 appear anywhere on the site
- Training on user input is on by default; opting out means knowing about a setting called "Improve the model for everyone"
- The pricing page warns of a significant increase to come, with neither a date nor an amount
- No postal address is published, and there is no about page, contact page or features page — support runs through email and an in-product form
- Liability is capped at USD 100 or twelve months of fees, whichever is greater, and account deletion is irreversible
- Only English and Chinese are documented as interface languages, and no robots.txt is served although the terms invoke the robots exclusion protocol
Pricing & Plans
A permanent free plan is available: the web assistant and the iOS and Android applications are free of charge, with no paid consumer tier. Charges apply to the API only, which is billed per million tokens against a prepaid balance. The lowest published entry point is USD 0.14 per million input tokens on a cache miss with deepseek-v4-flash. DeepSeek states on its pricing page that it plans a significant increase in the near future.
- DeepSeek Chat and DeepSeek App — free
- full access to the flagship model
- no paid tier
- API
- deepseek-v4-flash — USD 0.0028 per 1M cached input tokens
- USD 0.14 per 1M uncached input tokens
- USD 0.28 per 1M output tokens
- concurrency limit 2
- 500
- API
- deepseek-v4-pro — USD 0.003625 per 1M cached input tokens
- USD 0.435 per 1M uncached input tokens
- USD 0.87 per 1M output tokens
- concurrency limit 500
- Open-source models — weights and inference code under the MIT License
- free of charge
- with hosting at the user's own expense
Data, GDPR & hosting
A consolidated view of how DeepSeek handles your data.
GDPR overview
Implementation is real but partial. The privacy policy carries a dedicated European Economic Area, Switzerland and United Kingdom clause, with a table of legal bases per purpose — contract, consent, legitimate interests, legal obligation — and the full catalogue of data subject rights, including access, rectification, erasure, restriction, portability, withdrawal of consent, objection and the right to complain to a local supervisory authority. Prighter Group is named as the Article 27 representative for the EU and the UK, reachable at rep_deepseek@prighter.com. Rights are exercised by writing to privacy@deepseek.com, with a right of appeal if refused. What is missing matters too: the acronym GDPR never appears in the body text, no data processing agreement is offered, no subprocessor is named, and the transfer safeguard for data held in China is mentioned without being identified.
Who owns the data?
The terms split ownership cleanly. You keep whatever rights you hold in the inputs you submit, and DeepSeek assigns to you whatever rights it holds in the outputs it returns — with the caveat that outputs are not guaranteed to be unique and that the assignment does not reach another user's outputs. DeepSeek retains everything on its own side: software, model weights and parameters, interfaces and trademarks. The controller is Hangzhou DeepSeek Artificial Intelligence Co., Ltd. It shares personal data with service providers, with entities of its own corporate group, and with law enforcement, public authorities or rights holders where it believes in good faith that disclosure is necessary. It states it does not sell personal data, run targeted advertising or profile users.
Reuse rights
Reuse is permitted broadly and without asking. The terms state that inputs and outputs may be applied to a wide range of cases, naming personal use, academic research, derivative product development and the training of other models through distillation, provided the use is lawful and consistent with the terms. Three obligations come attached. Anyone publishing an output must first check it for accuracy and must label it clearly as AI-generated, and must not remove or circumvent the AI-content labels the platform adds. Any output feeding a decision with legal or material consequences for a person — credit, education, employment, housing, insurance, legal or medical — has to pass through human review. Finally, the DeepSeek name, logos and domain names may not be used without permission.
Data retention & training
Hosting summary
DeepSeek is unusually direct on this point. The privacy policy states, twice, that to provide the services the company directly collects, processes and stores personal data in the People's Republic of China, and warns separately that data may be held on a server outside the country where the user lives. The controller is Hangzhou DeepSeek Artificial Intelligence Co., Ltd., and the terms are governed by mainland Chinese law. Where transfers out of certain countries occur, the policy promises appropriate safeguards but names none — no standard contractual clauses, no adequacy decision, no binding corporate rules. Group entities handle storage, content delivery, security, research, model training and support; third-party providers are described only by category and never identified, and no subprocessor list, data processing agreement or security certification is published. The only third parties named anywhere are AWS and ishumei, in the cookie table. Note that the marketing site itself resolves to an Amazon anycast address in the United States: that is website delivery, not user-data hosting.
Things to keep in mind
Risks and trade-offs to weigh before adopting DeepSeek.
- Jurisdiction is the central risk: data is stored in China, the terms are governed by mainland Chinese law, and any dispute goes to the court covering the company's registered office
- The terms reserve the right to review user behaviour by technical means, including risk-filtering mechanisms and databases of illegal content features, and the privacy policy allows disclosure to law enforcement and public authorities
- Training on your inputs is the default state, so silence counts as consent — the opt-out only protects users who go looking for it
- Conversations shared through a unique link can be scraped by third parties; DeepSeek acknowledges the risk and says the robots exclusion protocol cannot fully prevent it
- The model can be confidently wrong: DeepSeek states hallucination cannot be fully avoided, requires human review before any decision with legal or material impact, and forbids removing its AI-content labels
- A free, unlimited and endlessly agreeable assistant invites over-reliance — outsourcing judgment, reasoning and writing until the underlying skill quietly erodes
- Recourse is minimal if something goes wrong: liability is capped at USD 100 or twelve months of fees, and deleting your account destroys your content irreversibly while DeepSeek may still retain data the law requires
Setup & Integrations
Technical difficulty
Negligible for everyday use: create an account with an email address or a third-party sign-in, then start typing on the web or in the mobile app. Low for developers: register on the platform, generate a key, change a base URL. Because the API accepts both the OpenAI and the Anthropic request formats, existing code usually needs no rewriting, and redirecting a coding agent takes two or three environment variables — Claude Code additionally wants Node.js 18 or later. Only self-hosting is genuinely demanding: the MIT-licensed weights are free, but the GPU infrastructure to run them is entirely the user's problem.
Deployment
Apps stores
Integrations
Supported languages
Behind DeepSeek
Fundraising
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement DeepSeek.
Frequently asked questions
Is DeepSeek free to use?
Where is my data stored?
Are my conversations used to train the models?
Who owns what I type and what the model answers?
Can I use the outputs to train another model?
Is there an API, and how compatible is it?
Is DeepSeek really open source?
Does the GDPR apply, and who represents DeepSeek in Europe?
What is the minimum age?
How do I reach support?
Should you pick DeepSeek?
DeepSeek offers one of the most aggressive capability-to-price ratios on the market, and pairs it with an openness rare at this level. The assistant is free on every platform with no upsell, the API starts at USD 0.14 per million input tokens, and the weights are published under the MIT License with technical reports and model cards to match. Because the API speaks both the OpenAI and the Anthropic dialects, trying it costs a base URL rather than a rewrite — which is why so many coding agents already list it as a supported backend.
The trade-off is stated plainly by DeepSeek itself: personal data is collected, processed and stored in the People's Republic of China, and the terms are governed by mainland Chinese law with jurisdiction at the company's Hangzhou registered office. Whether that is acceptable is a question of context rather than of quality, and it is the first thing a European or American organisation should settle. Around it sit narrower gaps: no data processing agreement, no named subprocessors, no security certification, no postal address, and a liability cap of USD 100.
Two details deserve attention before committing. Training on user input is enabled by default and switching it off requires finding a setting named Improve the model for everyone. And the pricing page announces a significant increase to come, without a date or a figure — so today's numbers should be treated as a snapshot rather than a plan.
For individuals and for teams whose data is not sensitive, DeepSeek is an unusually good deal, and its documentation is more forthcoming than most. For regulated workloads, the hosting jurisdiction and the missing contractual paperwork are the deciding factors, not the model.
- Choosing a selection results in a full page refresh.
- Opens in a new window.