Minute Verdict
DeepSeek is two products under one name. The chat is free once you make an account, with Google or Apple sign in, and it carries the DeepThink and Search modes. The API is sold per token and is cheap at the bottom of its range: $0.003 per million input tokens on flash when the cache is warm off-peak, up to $3.96 per million output tokens on v4-pro at peak.
Take the free chat if you want a general assistant without a subscription. Pay for the API only if you are building on it. If you do, run the bulk of your work off-peak, where every rate is exactly halved, and keep an eye on the cache: a repeated prompt at $0.003 against $0.15 for the same tokens uncached is a fifty-fold difference on the input side.
What DeepSeek does
DeepSeek is a Chinese AI company that publishes models and sells access to them through an API. Its site lists six generations of research, from V3 through V3.1, V3.2 and V4 to V4.1 and R1, and it runs four products: the web chat, a desktop harness, the API platform and the API documentation.
The company is registered in China, under Zhejiang registration numbers 浙ICP备2023025841号 and 浙B2-20250178, and its privacy policy states that personal data is stored in China. That is not an opinion about the tool, it is a sentence in the document, and it is the fact most buyers want before they decide.
The chat sits behind an account. The access page offers a login form, Google sign in and Apple sign in, and nothing else: there is no anonymous session and no guest mode. What you get inside is the familiar pair of modes the site advertises on its home page, DeepThink for reasoning and Search for web answers.
The API catalogue carries two models, and their names mislead. Flash sounds like the light one, but it is the model with vision, the higher concurrency limit of 2,500 requests, and the lower price. v4-pro is the stronger writer and carries no image input at all. Both take a 1M token context and return up to 384K tokens, which is more room than most rivals publish.
On this page
What you get for the money
The chat has no published price because there is nothing to price: an account is all it asks for. The API is metered, and the meter has four settings: which model, cache hit or miss, peak or off-peak, and input or output.
| Per million tokens | flash off-peak | flash peak | v4-pro off-peak | v4-pro peak |
|---|---|---|---|---|
| Input, cache hit | $0.003 | $0.006 | $0.022 | $0.044 |
| Input, cache miss | $0.15 | $0.30 | $0.66 | $1.32 |
| Output | $0.60 | $1.20 | $1.98 | $3.96 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is off-peak, weekends and holidays included, and off-peak is half price. In practice a European or American team works most of its day in the peak band, while a batch job scheduled at night halves the bill without changing a line of code.
Search and fetch are separate products billed per request, at $0.75 per thousand successful calls each. There is a deadline worth noting: tool calls made server-side through the Harness, Claude Code or Codex are free until November 1, 2026 at 00:00 Beijing time, and billed per request after that.
The billing order is stated plainly: granted balance is consumed before topped-up balance. If you are testing on credit, that means your own money stays untouched until the free balance runs out.
Where it falls short
The data question is the one that decides the purchase, and it is settled in the policy rather than in the marketing. DeepSeek is a Chinese company, its registration is Chinese, and its privacy policy states that personal data is stored in China. Anyone whose work involves client data, health information or anything covered by a contract should read that sentence twice before pasting a document into the free chat.
The naming of the two models will cost you money if you skim. Flash is cheaper, faster, supports images and allows 2,500 concurrent requests. Pro costs roughly three times more on output and cannot read an image at all. A team that assumes flash is the downgrade will overpay, and a team that assumes pro is the upgrade will discover the missing vision after wiring it up.
Peak hours land in the European and American working day. The cheap band is the one you use by accident, not the one you get by default, so the headline rate is not the rate most buyers will actually pay.
Two smaller points. The chat is behind an account, so there is no anonymous way to try it, and the sign in options are Google, Apple or an email registration. And the API documentation and the product site speak in two different rhythms, one technical and one promotional, which makes the model lineage harder to follow than the table suggests. Our how we test page explains what we do and do not measure.
DeepSeek alternatives
The three names below are what buyers compare against. Each line uses the DeepSeek API rate as the yardstick, because that is the figure you are trying to beat.
- ChatGPT, the reference assistant from OpenAI, sold as a monthly subscription rather than per token. Weigh it against the $0.60 per million output tokens here if you are deciding between a seat and a meter: AI chat category.
- Claude, aimed at long documents and code, also sold as a seat. Weigh it against the $3.96 per million output tokens of v4-pro at peak, which is the closest comparison in this catalogue: AI guides.
- Qwen, the other large Chinese family, distributed through several hosts rather than one platform, which changes the data question rather than the price. Weigh it against the $0.15 per million input tokens on flash: current deals.
None of the three publishes a per-token rate on a page reachable today in the same shape as this table, which is why the figures above are DeepSeek’s.
Who it is for: DeepSeek
DeepSeek suits two kinds of buyer. The first wants a capable general assistant, with reasoning and web search, without a monthly subscription, and does not mind creating an account for it. The second builds software and wants a cheap model with a very large context, where a 1M token window and a $0.15 input rate change what is financially possible.
It does not suit anyone who cannot send their data to China. That is not a technicality, it is the deciding constraint, and no feature on the page outweighs it. It also does not suit a team that needs vision on its strongest model, since v4-pro cannot read an image.
How we checked
This is a research review. We have not used DeepSeek, the chat or the API, so there is no hands-on score, no latency measurement and no output sample.
Every price on this page was read on October 30, 2026, on the API pricing page, which answered normally. The chat itself could not be exercised: chat.deepseek.com answers with a login page, and a login requires an account we do not have. We therefore say nothing about censorship, answer quality or the behaviour of DeepThink, because none of it can be verified without an account.
We did not verify the founders, the funding history, the January 2025 market episode, the published benchmarks or the American restrictions mentioned in our brief. None of those appear on the pages we could read, and we prefer a shorter review to a decorated one. The legal registration numbers and the storage location come from the site’s own footer and privacy policy.
The affiliate position is stated where it applies, and our editorial policy covers how a page like this is funded.
What to do next
Make a free account and use the chat for a week. If it holds up for your work, you have saved a subscription and the question is closed. If you are building, start on flash rather than pro, schedule batch jobs outside 01:00 to 10:00 UTC, and keep your prompts stable so the cache does the work: the difference between $0.003 and $0.15 per million input tokens is your prompt design, not your provider.
FAQ
Is DeepSeek free? The chat is free once you create an account, with Google or Apple sign in. The API is not free, it bills per token from $0.003 per million input tokens on flash with a warm cache off-peak.
How much does the DeepSeek API cost? On flash, $0.003 to $0.30 per million input tokens and $0.60 to $1.20 per million output tokens depending on cache and peak. On v4-pro, $0.022 to $1.32 input and $1.98 to $3.96 output.
What are the peak hours? 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour is off-peak, including weekends, and off-peak rates are half the peak rates.
Which model should I use? Flash for anything with images or high concurrency, since vision and 2,500 concurrent requests are there. v4-pro only if you need its output quality and can pay roughly three times the output rate.
Where is my data stored? The privacy policy places the company and its storage in China. That is stated in the policy itself, and it is the one fact worth weighing before you paste anything confidential.
Sources
Sources
- https://api-docs.deepseek.com/quick_start/pricing (checked October 30, 2026)
- https://deepseek.com (checked October 30, 2026)
- https://chat.deepseek.com (checked October 30, 2026)
- https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html (checked October 30, 2026)
Update log
October 30, 2026: first publication, research review, prices read on the API pricing page.
