DeepSeek Review
Reviewed by Carlos Gil · Founder & Editor, The Best Network
Updated 7 min readEditorially reviewed
Quick facts
- Starting price
- Free app; API from $0.15 per 1M input tokens
- Free plan
- Yes
- Best for
- Developers who want low API costs or self-hosted open weights, and individuals who want a free assistant for non-sensitive work
- Platforms
- Web (chat.deepseek.com), iOS, Android, API, open weights
BestStacked Score
Weighted average of 5 categories. Methodology below.
Quick answer
Is DeepSeek worth using in 2026?
For low-cost API work and self-hosting, often yes; for sensitive data, think carefully. The DeepSeek chat app is free, and DeepSeek lists no paid consumer plan. Its API runs DeepSeek-V4.1-Flash from $0.15 per million input tokens off-peak ($0.30 at peak) and DeepSeek-V4-Pro from $0.66, both with a 1M-token context window, and the model weights are released under the MIT license. DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China and may use your inputs for training unless you opt out.
- Price
- Free chat app. API: deepseek-flash $0.15/$0.60 and deepseek-v4-pro $0.66/$1.98 per 1M input/output tokens off-peak; peak hours cost double.
- Current models
- DeepSeek-V4.1-Flash (released September 10, 2026) and DeepSeek-V4-Pro, each with thinking and non-thinking modes.
- Open weights
- V4.1-Flash, V4-Pro and earlier models are on Hugging Face under the MIT license, so you can run them on your own infrastructure.
- Watch out for
- Data is stored in China, Chinese law governs the terms, and inputs can be used for training unless you turn off "Improve the model for everyone".
The Verdict
DeepSeek is the strongest option here if cost is the deciding factor. The chat app is free with no paid tier listed, the API charges a fraction of what most US labs charge per token, and the same models are published as MIT-licensed open weights that a company can host itself. The catch is governance, not price: DeepSeek's privacy policy says personal data is collected, processed and stored in the People's Republic of China, its terms are governed by Chinese law with disputes heard in Hangzhou, and inputs may be used for training unless you opt out. DeepSeek also doesn't publish business plans with admin controls. For personal, non-sensitive work or a self-hosted deployment it's good value; for client or regulated data, an assistant with business terms is the safer choice. We score it 8.3/10.
Best for
Developers optimizing API cost, teams that want to self-host open weights, and individuals who want a free assistant for non-sensitive tasks.
Skip if
You handle client, health, financial or other regulated data and need a vendor with business-plan privacy terms, admin controls and support.
The Background
What Is DeepSeek?
DeepSeek is an AI assistant and model family from Hangzhou DeepSeek Artificial Intelligence Co., Ltd., the company its privacy policy and terms name as the operator. The free chat app runs on the web at chat.deepseek.com and on iOS and Android, with a DeepThink button for step-by-step reasoning and a Search button for web results. Developers reach the same models through an API that accepts both OpenAI-format and Anthropic-format requests. As of October 12, 2026, the API lists two models: deepseek-flash, which maps to DeepSeek-V4.1-Flash, and deepseek-v4-pro, which maps to DeepSeek-V4-Pro-0813. DeepSeek also publishes its model weights on Hugging Face under the MIT license. The older deepseek-chat and deepseek-reasoner API names were retired on July 24, 2026, according to DeepSeek's changelog.
The Product
DeepSeek Features: What You Actually Get
Free chat app with DeepThink and Search
The web and mobile apps are free; DeepSeek's homepage describes the service as free to chat with, and no paid consumer plan is listed. DeepThink turns on step-by-step reasoning, and Search adds web results. DeepSeek's help centre covers image and file uploads, chat sharing and export.
A student uses DeepThink to work through a statistics problem step by step, then checks the method against the textbook.
V4.1-Flash and V4-Pro models
Both API models support a 1M-token context window, output of up to 384K tokens, and thinking and non-thinking modes, with thinking effort set to low, high or max. Image input (vision) is supported on deepseek-flash only. DeepSeek released V4 in April 2026, made V4-Pro generally available on August 13 and released V4.1-Flash on September 10.
A developer sends a long codebase excerpt to deepseek-v4-pro in thinking mode to plan a refactor, then uses deepseek-flash for routine completions.
Low-cost API with off-peak pricing
deepseek-flash costs $0.15 per million cache-miss input tokens and $0.60 per million output tokens off-peak, and deepseek-v4-pro $0.66 and $1.98. Peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays) cost double. Cached input is billed at a fraction of the cache-miss rate, and context caching is automatic.
A startup schedules its nightly batch summarization jobs outside DeepSeek's peak hours to pay the off-peak rate.
OpenAI- and Anthropic-compatible API
The API accepts OpenAI-format requests at api.deepseek.com and Anthropic-format requests at api.deepseek.com/anthropic, and supports JSON output, tool calls and a Responses API. Chat prefix completion and fill-in-the-middle completion are in beta.
A team points an existing OpenAI SDK integration at DeepSeek's base URL to compare cost on the same prompts.
MIT-licensed open weights
DeepSeek publishes V4.1-Flash, V4-Pro, V4-Flash and V3.2 weights on Hugging Face under the MIT license. Self-hosting means prompts never reach DeepSeek's servers, though running a model of this size needs serious GPU infrastructure. DeepSeek's model disclosure says all weights are MIT, but its OCR-2 model is Apache-2.0 and the V4 Base repositories carry no license tag.
A company with data-residency rules runs DeepSeek weights on its own cloud account instead of calling DeepSeek's API.
DeepSeek Harness (preview)
DeepSeek Harness is an open-source (MIT) desktop agent in public preview for macOS 13+ and Windows 10+, with Standard, Code, Minimal and Creator modes, voice input and scheduled tasks.
A developer tries Harness in Code mode on a side project to see whether it can replace a paid coding agent.
Feature recap
- Free web and mobile chat app with DeepThink and Search
- DeepSeek-V4.1-Flash and V4-Pro with a 1M-token context window
- API from $0.15 per 1M input tokens off-peak; peak hours cost double
- OpenAI- and Anthropic-format API, JSON output and tool calls
- MIT-licensed open weights on Hugging Face
- DeepSeek Harness desktop agent in public preview
The Numbers
DeepSeek Pricing: Every Plan Explained
DeepSeek's chat app is free, and DeepSeek lists no paid consumer plan. The API is prepaid: you top up a balance by PayPal, bank card, Alipay or WeChat Pay, top-ups don't expire, and DeepSeek's FAQ says unused balance can be refunded. Prices below are per million tokens, checked on DeepSeek's pricing page on October 12, 2026. Off-peak rates are half the peak rates; peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.
| Plan | Monthly | Annual | Included Features | Best For |
|---|---|---|---|---|
| Chat app (web, iOS, Android)Best value | Free | Free | DeepThink reasoning, Search, file and image uploads, chat sharing and export. No paid consumer plan listed | Individuals and non-sensitive work |
| API: deepseek-flash (V4.1-Flash) | Off-peak: $0.15 input (cache miss), $0.003 cached, $0.60 output | Peak: $0.30 input, $0.006 cached, $1.20 output | 1M context, up to 384K output, thinking and non-thinking modes, vision, JSON output, tool calls | High-volume and cost-sensitive API work |
| API: deepseek-v4-pro (V4-Pro-0813) | Off-peak: $0.66 input (cache miss), $0.022 cached, $1.98 output | Peak: $1.32 input, $0.044 cached, $3.96 output | 1M context, up to 384K output, thinking and non-thinking modes, no vision | Harder reasoning and coding tasks |
| Open weights | Free to download (MIT license) | You pay for your own hosting | V4.1-Flash, V4-Pro, V4-Flash and V3.2 weights on Hugging Face | Teams that need to keep data on their own infrastructure |
For API users the math is simple: deepseek-flash costs $0.15 per million cache-miss input tokens off-peak, so a job that sends 100 million input tokens and gets 20 million output tokens back costs about $15 + $12 = $27 off-peak, or $54 at peak, before any cache discount. That is our calculation from DeepSeek's published prices, not a quote. In the table above, the Monthly column holds off-peak rates and the Annual column peak rates, because DeepSeek has no subscription plans. Note that DeepSeek mentions a granted balance and a free quota in its pricing page and platform terms without giving an amount. For the chat app there is nothing to buy, so the real cost is the data terms covered below.
Read before you sign up
Where DeepSeek stores your data, and how it's used
Location: DeepSeek's privacy policy (last updated February 10, 2026) says: "we directly collect, process and store your Personal Data in People's Republic of China." The operator and data controller is Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
What's collected: account details (such as email or phone number and date of birth), everything you enter, including text, voice, files, photos and chat history, plus device identifiers, IP address and approximate location.
Training: the privacy policy says DeepSeek uses data to train and improve its models and lists a right to opt out. The Terms of Use (last updated March 27, 2026, section 4.3) say de-identified inputs and outputs may be used and name the setting: turn off "Improve the model for everyone". The privacy policy itself doesn't say how to opt out. We found no official page saying whether API data is treated differently.
Sharing: the policy says entities in DeepSeek's corporate group support the service, including for model training, and that data may be shared with law enforcement and public authorities where required by law.
Retention and deletion: data is kept as long as your account is active. Deleting your account is permanent, and DeepSeek's FAQ says any remaining developer balance is forfeited.
Law and courts: the Terms of Use and the Open Platform Terms (effective April 29, 2026) are governed by the law of mainland China, with disputes going to the court where Hangzhou DeepSeek's registered office is.
Age: the privacy policy says the service isn't aimed at children, the Terms of Use say it's primarily for adults and under-18s need a guardian's consent, and the policy's EEA section mentions users aged 14 to 17.
Read before you buy
Where DeepSeek's own pages disagree
V4-Pro routing: DeepSeek's September 10 announcement says that from September 14 all V4-Pro requests route to V4.1-Flash at Flash prices. Its changelog says V4-Pro continues after September 14 with billing unchanged, and the pricing page still lists deepseek-v4-pro at Pro prices. We report the pricing page and changelog, and note the conflict.
Licensing: DeepSeek's model disclosure says all model weights are released under the MIT license, but DeepSeek-OCR-2 is Apache-2.0 on Hugging Face and the V4 Base repositories show no license tag.
Training opt-out: the privacy policy lists the right to opt out without saying how; only the Terms of Use name the setting.
The Tradeoffs
DeepSeek Pros and Cons
What we love
- Free chat app with reasoning (DeepThink) and web search
- API prices well below most US labs, halved off-peak
- 1M-token context window on both API models
- MIT-licensed open weights you can host yourself
- OpenAI- and Anthropic-compatible API makes switching easy
- Prepaid API balances don't expire and can be refunded, per DeepSeek's FAQ
What could be better
- Personal data is collected, processed and stored in China
- Terms governed by Chinese law, with disputes heard in Hangzhou
- Inputs can be used for training unless you opt out
- No business plan with admin controls, SSO or a data-processing agreement listed
- DeepSeek's pages disagree on how V4-Pro requests are routed
- Some email domains and IP addresses can't register, per DeepSeek's FAQ
The Audience
Who Should Use DeepSeek?
Best for: developers optimizing API cost
If you run high token volumes, deepseek-flash at $0.15 per million input tokens off-peak and OpenAI-compatible endpoints make DeepSeek cheap to test against your current provider.
Best for: teams that want to self-host
MIT-licensed weights let you run DeepSeek models on your own infrastructure, so prompts never reach DeepSeek's servers.
Skip if: you handle client or regulated data
DeepSeek stores personal data in China and lists no business plan with admin controls. ChatGPT Business, Claude Team or Gemini in Workspace offer business terms instead.
The Methodology
How We Reviewed DeepSeek
This is an editorial review, not a documented hands-on test. We didn't use a DeepSeek account for this review. It's based on DeepSeek's pricing page, API changelog, news posts and guides, its homepage and help-centre FAQ, its Privacy Policy, Terms of Use, Open Platform Terms and model disclosure, and its Hugging Face and GitHub pages, all checked on October 12, 2026. We didn't sign in, so we couldn't check what the logged-in app shows. The cost example is our calculation from published prices. Where DeepSeek's pages conflict, as on V4-Pro routing, we report both statements. No benchmark results are claimed.
Our editorial evaluation considers product information, official documentation, pricing, features and professional experience. It is not a claim of documented hands-on testing.
Our review methodology →What changed
- : Expanded from a short research preview into a full editorial review. Checked on DeepSeek's pricing page, changelog, news posts, FAQ, Privacy Policy (February 10, 2026), Terms of Use (March 27, 2026), Open Platform Terms and Hugging Face on October 12, 2026. Replaced the preview's retired model names with V4.1-Flash and V4-Pro and added peak and off-peak API prices, the data-location and training terms, and conflicting-page notes. Overall score kept at the preview's 8.3; sub-scores are provisional.
The Field
DeepSeek vs The Competition
DeepSeek competes with general AI assistants on the consumer side and with frontier-model APIs on the developer side. Scores come from each tool's own BestStacked review.
| Tool | Starting Price | Free plan | Best Feature | BS Score |
|---|---|---|---|---|
| DeepSeek | Free app; API from $0.15 per 1M input tokens | Yes | Free chat app with DeepThink and Search | 8.3/10 |
| ChatGPT | Free; Go $8/mo; Plus $20/mo | Yes | Broadest feature set in one app | 9.1/10 |
| Claude | Free; Pro $20/mo ($17/mo billed yearly) | Yes | Long-form writing and long-context analysis | 9/10 |
| Gemini | Free; Google AI Pro $19.99/mo | Yes | Works inside Gmail, Docs and Drive | 8.9/10 |
ChatGPT: ChatGPT covers more jobs in one app, including image generation and voice, and its Business plan keeps business data out of training by default. DeepSeek is free and far cheaper on the API.Read our ChatGPT review →See DeepSeek vs ChatGPT →
Claude: Claude suits long documents and careful writing and offers Team and Enterprise plans with admin controls. DeepSeek wins on price and open weights.Read our Claude review →
Gemini: Gemini also has a broad free tier and adds Workspace integration and business terms through Google Workspace. DeepSeek offers open weights and lower API prices.Read our Gemini review →
In The Wild
How People Actually Use DeepSeek
A developer building a summarization feature
Prototype on deepseek-flash through the OpenAI-compatible endpoint, run batch jobs off-peak, and compare output and cost against your current provider before committing.
A company with strict data-residency rules
Skip DeepSeek's hosted API and app, and evaluate the MIT-licensed weights on your own cloud or hardware instead.
A student or hobbyist
Use the free app with DeepThink for maths and coding practice, keep personal details out of prompts, and turn off "Improve the model for everyone" if you don't want your chats used for training.
The Ecosystem
DeepSeek Integrations
DeepSeek's integration story is its API compatibility: tools that speak the OpenAI or Anthropic API formats can usually point at DeepSeek by changing the base URL and key.
- →OpenAI-format API (api.deepseek.com)
- →Anthropic-format API (api.deepseek.com/anthropic)
- →Responses API, JSON output and tool calls
- →Hugging Face (open weights)
- →GitHub (deepseek-ai repositories, including DeepSeek Harness)
- →Web app, iOS and Android
DeepSeek doesn't list native connectors to Google Workspace, Microsoft 365, Slack or CRMs, so integrations depend on third-party tools that support its API. Check that any tool you connect is comfortable sending your data to DeepSeek's servers in China.
Common Questions
DeepSeek FAQ
The Final Word
Should You Use DeepSeek?
DeepSeek is the clearest value pick among AI assistants: a free chat app, API prices well below most US labs, a 1M-token context window and MIT-licensed open weights. Whether that value is worth it depends on your data. DeepSeek's own privacy policy says personal data is stored in China, its terms fall under Chinese law, and chats can be used for training unless you opt out. Use the app for non-sensitive personal work, use the API or open weights where cost matters and the data allows it, and choose an assistant with business terms for client work. We score DeepSeek 8.3/10.
Yes, if
You want a free assistant for non-sensitive work, the lowest API cost, or open weights you can host yourself.
No, if
You handle client or regulated data, need admin controls and support, or can't send data to servers in China.
About the reviewer
Carlos Gil
Founder & Editor, The Best Network
Carlos Gil is the bestselling author of The End of Marketing, a LinkedIn Top Voice, and founder of The Best Network. He brings 15+ years in marketing to practical reviews of AI, SaaS, email marketing, creator, and ecommerce software.
Get the best of BestStacked
The AI tools worth knowing. None of the noise.
- New AI tool and software reviews as they publish
- Head-to-head comparisons and stack breakdowns
- The occasional recommendation worth knowing
No fake urgency. No made-up weekly cadence. Unsubscribe anytime.