DeepSeek

DeepSeek is a Chinese AI lab, operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., that sells its own language models in two ways: a developer API and a chat assistant on the web and in apps. The API serves two models, deepseek-flash and deepseek-v4-pro, both with a 1M-token context window and a maximum output of 384K tokens.

Developers can call it with OpenAI-style or Anthropic-style requests by changing the base URL, so existing SDK code needs little rewriting. The vendor also documents integrations with coding agents and offers separate Search and Fetch APIs. The chat product shares a privacy policy, last updated 10 February 2026, that says personal data is processed and stored in the People's Republic of China. Billing for the API is per million tokens, prepaid from a topped-up balance.

• OpenAI and Anthropic API formats. The same models answer at api.deepseek.com for OpenAI-format calls and at a separate /anthropic path for Anthropic-format calls. The docs say the official SDKs or compatible software work after a base URL and key change.
• Two models with thinking mode. deepseek-flash and deepseek-v4-pro both support thinking and non-thinking modes. The pricing page lists JSON output, tool calls and a Responses API for both, with vision only on Flash and FIM completion in non-thinking mode.
• Cache-aware token pricing. Input tokens are billed at a lower rate on a cache hit than a cache miss. For Flash the page lists $0.003 per million cached input tokens off-peak, against $0.15 for a miss, so repeated prompts cost much less.
• Peak and off-peak rates. Off-peak rates are half of peak. The vendor defines peak as 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Batch jobs scheduled outside those hours cost half.
• Search and Fetch API. A newer standalone API gives web search and page fetch from the vendor's own search engine. It is billed per request, $0.75 per 1,000 successful requests for each, separate from model tokens.
• Agent tool integrations. Docs pages cover using DeepSeek as the backend for Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix, WorkBuddy/CodeBuddy and Qoder. A DeepSeek Harness for agent builders is in developer preview, says the vendor.

• A startup with an existing OpenAI integration. Points its current SDK at the DeepSeek base URL and swaps the model name to test lower per-token cost, without rewriting request code. Compliance review of where data goes comes first, since the vendor processes data in China.
• A team running overnight batch jobs. Schedules summarisation or classification outside the weekday peak windows to pay the off-peak rate, which the vendor sets at half the peak price, and reuses long prompts to benefit from the cache-hit rate.
• A solo developer using a coding agent. Configures an agent tool such as Claude Code or OpenCode to use DeepSeek as its model backend, following the vendor's agent integration guides, and pays only for tokens used from a topped-up balance.

Create an account on the DeepSeek Platform, apply for an API key, and top up a balance; the docs state that fees are deducted from the topped-up or granted balance, with granted balance used first. Then set base_url to https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format), pass the key as a bearer token and choose deepseek-flash or deepseek-v4-pro.

The Your First API Call page has curl, Python and Node.js examples, and a thinking parameter turns reasoning on or off. The chat assistant needs no setup beyond an account on the website or app. The docs also list rate limits and error codes on separate pages.

Hangzhou DeepSeek Artificial Intelligence runs the web and app assistant alongside its API. The privacy policy says personal data is collected, processed and stored in the People's Republic of China, and terms are governed by mainland Chinese law.

Users can copy or delete chat history in settings. Its site shows no consumer plan table, so the cost of the chat product is unknown to us.

In our AI chatbots and assistants ranking DeepSeek stands at #6 of 14: it follows Poe and comes before Kimi. Challengers hold places 1 to 3 where a category has them; the number here is a place in the order, not a mark out of ten.

Counting headquarters, 2 out of 14 sit in Asia-Pacific; for DeepSeek that is China.

Pricing as recorded: Per-token API published; chat plans not shown on its site.

API prices are per million tokens in US dollars. Flash: input cache hit $0.003 off-peak and $0.006 peak; cache miss $0.15 off-peak and $0.30 peak; output $0.60 off-peak and $1.20 peak. V4-Pro: cache hit $0.022 and $0.044; cache miss $0.66 and $1.32; output $1.98 and $3.96. Concurrency limits are listed as 2500 for Flash and 500 for V4-Pro. Search and Fetch cost $0.75 per 1,000 successful requests each. The vendor reserves the right to change prices. No subscription plans for the chat assistant were published on the pages we read.

Source: the pricing material on deepseek.com, as read for Best AI Chatbots and Assistants in 2026 and last reviewed on 6 November 2026.

The documentation gives separate base URLs for OpenAI and Anthropic clients, so existing SDK code can usually be repointed.

The privacy question is the real filter: the terms reference de-identified processing but no location, and the governing law is that of the People's Republic of China.

The privacy policy says DeepSeek directly collects, processes and stores personal data in the People's Republic of China and that the controller is Hangzhou DeepSeek Artificial Intelligence Co., Ltd.

Users can manage chat history in settings, including copying or deleting it, and the policy lists rights such as deletion and data transfer where applicable law gives them. It says the services are not aimed at children. Security certifications such as SOC 2 or ISO 27001 were not stated on the pages we read.

Integrations the vendor names (12):

• OpenAI SDK
• Anthropic SDK
• Claude Code
• Codex
• OpenCode
• OpenClaw
• Hermes
• Reasonix
• WorkBuddy/CodeBuddy
• Qoder
• GitHub Copilot
• DeepSeek Harness (preview)

Also in our records:

• The documentation gives separate base URLs for OpenAI and Anthropic clients, so existing SDK code can usually be repointed.

We pick the closest products in the same guide by tags, price model and region.

• ChatGPT (#4 of 14): Suits: Organisations that want one vendor with plans from solo use to enterprise.
• Kimi (#7 of 14): Not for: teams needing English-first admin documentation and clear prices. Moonshot AI assistant with membership tiers and a business edition.
• Microsoft Copilot (#10 of 14): Not for: teams that want a standalone chat tool without Office. Microsoft assistant sold inside Microsoft 365 plans and as a business add-on.

Read from the vendor’s own pages on 5 November 2026.

• api-docs.deepseek.com/quick_start/pricing: Models, prices, peak hours, concurrency, Search and Fetch pricing, deduction rules
• api-docs.deepseek.com: Base URLs, formats, first API call, agent integrations
• cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html: Controller, storage in the PRC, chat history, rights

The same four questions are asked of every AI chatbots and assistants product: time to value, real cost, the way out and the intended buyer. That puts DeepSeek at #6 of 14. Method and commercial relationships are in our editorial process.