AI assistantsDeepSeek at a glance
Pros
- Free web and mobile chat with no paid tier needed for everyday use
- API prices far below most Western frontier models, with half-price off-peak hours
- Cache hits on repeated prompt prefixes cost almost nothing
- OpenAI-compatible and Anthropic-format endpoints make switching simple
- Open model weights let you self-host and keep data in-house
Cons
- Data is processed and stored in China under DeepSeek's privacy policy
- US federal civilian agencies must remove DeepSeek products, and many organisations block it
- Model names and prices changed several times in 2026, so code and budgets need maintenance
- Per-account concurrency limits can cause throttling at scale
- Answer quality can be inconsistent on niche or very recent topics
What DeepSeek is
DeepSeek is an AI research company based in Hangzhou, China, that builds large language models and offers them in three ways: a free chat assistant on the web and mobile, a pay-as-you-go developer API, and open model weights that anyone can download and run. It became widely known in early 2025, when its reasoning model matched much more expensive US models on many benchmarks at a far lower price.
In 2026 the line-up changed quickly. DeepSeek V4 launched in April, the old deepseek-chat and deepseek-reasoner API names were retired in July, and by November the API offered two models: deepseek-flash, now running V4.1-Flash, and the larger deepseek-v4-pro. Both handle a context window of up to one million tokens and run in a thinking mode by default.
Who DeepSeek is for
The strongest case for DeepSeek is developers and startups that need a capable reasoning and coding model inside their own product and care about cost per task. Workloads with long, repeated prompts, such as support bots with fixed policies or document pipelines, benefit most from its very cheap cache-hit pricing.
Individuals who want a free, capable chatbot for general questions, writing and coding help will also find the app useful. It is a poor fit for teams handling confidential client files, health or financial records, or anything subject to government security rules, because of where the data is stored and the growing list of organisations that restrict it.
Key features
The chat app covers the basics of a modern assistant: conversation history, file and image uploads, web search and a thinking mode that shows its reasoning before it answers. Users on G2 like the clean interface and small touches such as saving an image of a whole conversation, though there are no team workspaces, shared projects or admin controls of the kind ChatGPT and Claude sell to businesses.
On the API, deepseek-flash is the default model. It supports image input, JSON output and tool calls, and DeepSeek describes the October 2026 update as faster and less verbose than the version before. deepseek-v4-pro is the larger model, reported as 1.6 trillion total parameters, aimed at harder coding, maths and multi-step reasoning, but it does not accept images.
Both models expose thinking and non-thinking modes and a reasoning-effort setting of low, high or max. The reasoning text comes back in a separate field from the answer, which makes it easy to log or hide. The API speaks the OpenAI Chat Completions and Responses formats and also offers an Anthropic-format endpoint, so most frameworks work by changing the base URL, key and model name.
DeepSeek also publishes model weights on Hugging Face and GitHub. That matters for privacy-sensitive buyers: a model run on your own servers, or by a hosting provider you choose, does not send prompts to DeepSeek. Older R1 and V3 weights remain available for self-hosting even though those models have left the official API.
Pricing explained
For individuals, DeepSeek costs nothing. The chat app has no paid tier, which is its biggest advantage over ChatGPT Plus, Claude Pro or Gemini's paid plans for people who only need a capable general assistant.
For developers, the numbers are the story. At peak rates deepseek-flash costs $0.30 per million input tokens and $1.20 per million output tokens, and deepseek-v4-pro costs $1.32 and $3.96. Off-peak, both halve. A worked example in the November guide puts 1,000 requests with 2,000 input tokens and 500 output tokens each at well under a dollar on Flash, because most of the input hits the cache.
The cheapest small models from OpenAI, Anthropic and Google can undercut Flash on list price, so DeepSeek is not automatically the lowest bill. Its edge is getting a large, capable reasoning model for small-model money. Test on your own prompts and measure cost per successful task rather than cost per token.
Ease of use and setup
The chat app needs only an account and works much like any other assistant. Getting started with the API is equally simple: create an account, add a balance, generate a key and point an OpenAI SDK at DeepSeek's base URL.
The harder part is keeping up. In 2026 alone DeepSeek retired two model IDs, introduced peak and off-peak pricing, launched two new models and redirected legacy names to new versions. Code that still called deepseek-chat stopped working in July. Teams should pin model names, watch the change log and keep a fallback provider ready.
Integrations
Because the API is OpenAI-compatible, DeepSeek drops into LangChain, LlamaIndex, the Vercel AI SDK and most coding tools that let you set a custom endpoint. The Anthropic-format endpoint extends that to tools built for Claude.
You can also reach DeepSeek models through routers such as OpenRouter and inference hosts such as Together AI and Fireworks. Their prices and model versions differ from DeepSeek's own, but they can offer other regions, failover and consolidated billing. The consumer app itself has no business connectors for drives, email or calendars.
Privacy, security and reliability
DeepSeek's privacy policy, updated in February 2026, says it collects prompts, chat history, uploads and device data, and processes and stores personal data in the People's Republic of China. Users can delete chat history and opt out of training, but the data still falls under Chinese law.
Governments have reacted. In October 2026 the US Department of Homeland Security published a notice requiring federal civilian agencies to remove DeepSeek products, including some third-party services built on its technology, and banning future purchases. It is not a consumer ban, but many companies follow similar rules. Regulators in several other countries have investigated or restricted the app since 2025.
Operationally, the API limits concurrent requests per account, 2,500 for Flash and 500 for Pro, and returns errors above that. You can apply for more at no extra cost. DeepSeek does not publish an uptime SLA, and users on review sites mention slow responses on heavy reasoning tasks.
What users like and dislike
Reviewers consistently praise the price, the quality of reasoning and coding answers for the money, and the simple interface. Developers value the cache pricing and OpenAI compatibility, which make switching from another provider quick.
Complaints focus on inconsistency: some answers are sharp, others generic or wrong on specialised and recent topics, so users rephrase. Others find it slower than rivals on complex requests. Privacy is the recurring concern in press coverage, and the pace of change frustrates teams that built on retired model names.
Our verdict
DeepSeek offers some of the best value in AI. The free app is a capable everyday assistant, and the API gives developers a large reasoning model at prices that make high-volume use affordable, with off-peak and cache discounts that reward good engineering.
The trade-off is trust. Data stored in China, a US federal removal order and few enterprise assurances mean it should not touch sensitive information through DeepSeek's own services. Use it for public or low-risk data, or self-host the open weights, and pick a Western provider with business contracts when confidentiality matters.
DeepSeek scores 7.7 out of 10 overall. Its strongest criterion is value for money (9.5/10) and its weakest is support and trust (5.5/10).
DeepSeek pricing
| Plan | Price | What you get |
|---|---|---|
| Chat app | $0 | Free web, iOS and Android chat with the current DeepSeek models |
| API: deepseek-flash | $0.30 / $1.20 per 1M tokens | Input (cache miss) / output at peak; cache hits $0.006 per 1M; half price off-peak |
| API: deepseek-v4-pro | $1.32 / $3.96 per 1M tokens | Input (cache miss) / output at peak; cache hits $0.044 per 1M; half price off-peak |
| Search and Fetch APIs | $0.75 per 1,000 requests | Standalone web search and page fetch, billed per successful call |
| Self-hosted weights | Free to download | Run open DeepSeek models on your own hardware; you pay for compute |
DeepSeek's API is prepaid and pay-as-you-go. You top up a balance on the developer platform and each request is deducted as tokens multiplied by the model's price. There is no monthly subscription and no seat fee, and promotional credit, when you have any, is used before your topped-up balance.
We could not read DeepSeek's own pricing page directly, so the figures above come from a developer guide updated on 9 November 2026 that quotes the official Models and Pricing page. Another November 2026 price tracker gives the same $0.30 to $3.96 range. DeepSeek says it may change prices, so confirm before you budget.
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else, including weekends, is billed at half price, which makes batch jobs scheduled overnight in the US noticeably cheaper.
Output tokens cost three to four times as much as fresh input, and both models think by default, so reasoning text adds to the bill. Capping output length, turning thinking off for simple tasks and keeping system prompts identical between calls (to earn the cache-hit rate) are the main ways to control cost.
Prices in US dollars, checked on Nov 21, 2026. Confirm on DeepSeek's own site before you buy.

