DeepSeek review: Pay-per-token language model API with a web version, billed in 元
Pay-per-token language model API with a web version, billed in 元
Not published, free plan: no, prices checked on October 30, 2026.
Last verified · Checked against the official pricing page · How we verifySuggest an update
Written and checked by AI agents from the vendor’s official pages. How we verify
What DeepSeek does
Provides language models through a pay-per-token API with OpenAI- and Anthropic-format endpoints, plus Search and Fetch APIs for web search and page retrieval.
DeepSeek is the product name used by a Chinese AI research company. Its homepage says the company open-sources frontier models such as DeepSeek-V4 and DeepSeek-R1. The homepage also links to an app download, a web version, a Harness entry and an API open platform, without describing them further.
On the API side, the pricing page lists two models, deepseek-flash and deepseek-v4-pro. The table shows a single value for both models for context length (1M), maximum output (384K) and thinking mode, where thinking and non-thinking modes are supported and thinking is the default. Tool calls are marked as supported for both. Image understanding is marked as supported only for deepseek-flash. The API can be reached through an OpenAI-format or an Anthropic-format base URL. Separate Search and Fetch APIs cover web search and page retrieval.
Pricing not published on the official site as fixed plans: no monthly or yearly subscription is listed, and the pricing page shows only per-token rates. Charges are counted per million tokens in 元 and taken from a top-up balance or a gifted balance, with the gifted one used first. Off-peak prices are half of peak prices. For example, off-peak output costs 4元 per million tokens on deepseek-flash and 13.5元 on deepseek-v4-pro. Search and Fetch each cost 5元 per 1,000 successful requests. Concurrency is limited to 2500 for deepseek-flash and 500 for deepseek-v4-pro. The vendor reserves the right to change prices.
- 1M context length
- Maximum output of 384K
- Tool calls
- Thinking and non-thinking modes
- Image understanding on deepseek-flash
- OpenAI-format and Anthropic-format API
- Search and Fetch APIs for web search and page retrieval
Good for / Not for
Good for
- Developers who want to call models from code and pay per token
- Teams that need long context (1M) and long outputs (384K)
- Projects that need web search and page fetching through an API
Not for
- Buyers who want a fixed monthly subscription: no plans are listed
- Anyone who needs a stated free plan or trial: none appears on the pricing page
How to start
Limits and gotchas
- Concurrency is limited to 2500 for deepseek-flash and 500 for deepseek-v4-pro. source
- Peak-time prices are double off-peak prices; peak hours are weekdays in Beijing time. source
- Search and Fetch calls via Harness, Claude Code or Codex are charged from 2026-11-01 00:00 Beijing time. source
- The vendor may change prices, so check the pricing page regularly. source
From the official site and docs of DeepSeek, checked on October 30, 2026.
Pros and cons
Pros
- Usage-based billing per million tokens, with no subscription required
- Off-peak prices are half of peak prices
- Two API models to choose from, both listed with a 1M context length
- OpenAI-format and Anthropic-format base URLs are both offered
- Homepage says models such as DeepSeek-V4 and DeepSeek-R1 are open-sourced
Cons
- Pricing not published on the official site as fixed plans; only per-token rates appear
- No free plan or trial is stated on the pricing page
- Prices are shown only in 元 and the pricing page is in Chinese
- Concurrency is capped at 2500 for deepseek-flash and 500 for deepseek-v4-pro
- Search and Fetch are billed separately at 5元 per 1,000 successful requests each
Our take
DeepSeek suits developers and teams who want to call language models from their own code and pay only for the tokens they use, choosing between a flash model and a pro model. The listed context and output sizes point to long inputs and long answers. It is a weaker fit for buyers who want a fixed monthly subscription: Pricing not published on the official site as plans, and the pages state no free plan or trial. Costs depend on cache hits and on peak versus off-peak timing, so budgeting needs some work. The pricing page is in Chinese and quotes only 元. Search and Fetch charges for calls through Harness, Claude Code or Codex start on 2026-11-01, so check the page for current terms before relying on them.
Comparisons and lists with DeepSeek
Coming soon
Coming soon — we are writing comparisons that include DeepSeek.
DeepSeek pricing
Pricing not published on the official site. Check DeepSeek’s website.
Free trial: none listed. Free plan: no.
Prices checked on on the official DeepSeek pricing page. “—” = not published.
Frequently asked questions
How is the DeepSeek API priced?
By tokens, per million, in 元. Fees are deducted from a top-up balance or a gifted balance, gifted first. Off-peak prices are half of peak prices. No subscription plans are listed.
Which models does the API offer?
The pricing page lists deepseek-flash and deepseek-v4-pro. The table shows thinking and non-thinking modes, a 1M context length and a maximum output of 384K as one value for both, and tool calls as supported by both.
Does DeepSeek have a free plan?
The pricing page does not state a free plan or a trial.
What do web search and fetch cost?
Search and Fetch are separate APIs, each billed at 5元 per 1,000 successful requests. Calls made through DeepSeek Harness, Claude Code or Codex are not billed yet and will be charged from 2026-11-01 00:00 Beijing time.
Are there usage limits?
Concurrency limits are 2500 for deepseek-flash and 500 for deepseek-v4-pro.
Recently verified in AI Chatbots & Assistants
Gleap
Microsoft Copilot
Continue
FlowithAdd an AGIPicks badge to your site
The badge shows DeepSeek’s verification status on AGIPicks and links to this page. Light/dark and compact versions: badges.
Suggest an update
Wrong price, outdated fact or a tool that shut down? Are you the vendor and disagree with this page? Tell us, ideally with a link to the official page. We check every suggestion against the official pages within 7 days; a suggestion never changes the page by itself.