The full review
What is DeepSeek?
DeepSeek is an AI lab that ships the same model family through two very different front doors. On the consumer side there is DeepSeek Web and the DeepSeek App, which need no setup at all — the homepage, headlined Into the Unknown, drops you straight into a Chat Now button. On the developer side there is the API Platform, a hosted inference endpoint billed per token.
The current generation is DeepSeek V4, with DeepSeek-V4-Pro as the flagship, and the site still lists the earlier V3.2, V3.1, R1 and V3 releases. The vendor reports that the newest models are fully available across the web, mobile app, and API, so what you test in the chat window is broadly what you get from the endpoint.
The DeepSeek V4 model line
The API documentation lists three models: deepseek-v4-flash, deepseek-v4-pro, and an experimental deepseek-v4-flash-vision-exp. All three carry a 1M token context window with a maximum output of 384K tokens, which is real headroom for long contracts, sprawling codebases, or full meeting transcripts you would otherwise have to chunk and stitch back together. The vision variant accepts image input but is explicitly flagged experimental, so treat it as a preview rather than a production dependency.
Working with the DeepSeek API
The compatibility story is the strongest practical argument here. DeepSeek states its API uses a format compatible with OpenAI and Anthropic, so an existing integration usually moves over by swapping a base URL and a key rather than rewriting client code. The docs name Claude Code, GitHub Copilot, and OpenCode as tools that already run against the endpoint.
Feature coverage is broad enough for serious builds: Thinking Mode, Tool Calls, JSON Output, Multi-round Conversation, a Files API, plus beta Chat Prefix Completion and FIM Completion for code infilling. A Responses API sits alongside the Anthropic-compatible interface. DeepSeek even publishes recommended temperature values by task, from 0.0 for coding and math up to 1.5 for creative writing.
Context caching and off-peak billing
Two mechanics do most of the cost work, and both reward how you architect rather than what you negotiate. Context Caching on Disk is enabled by default for every user, with no code changes required. Each request builds a disk cache; when a later request shares a prefix with an earlier one, the overlap is served from that cache and billed at a far lower input rate. The API returns cache hit token counts separately, so the saving is measurable rather than theoretical.
The second mechanic is time of day. DeepSeek runs peak and off-peak rates, with off-peak set at half of peak, and the docs define peak hours as 01:00 to 04:00 and 06:00 to 10:00 UTC. Batch evaluation runs, bulk classification, and overnight summarisation all get cheaper for simply being scheduled outside those two windows.
Agent capability and DeepSeek Harness
The homepage leads on significantly enhanced agent capabilities for the latest release, along with support for the Responses API and Codex integration. DeepSeek Harness, the company's own agent tooling layer, appears in the site navigation, and the API documentation states that it is now in developer preview for agent harness developers worldwide. That status is worth respecting: it signals direction, not a finished product to build a roadmap around this quarter.
Who should choose DeepSeek
Two groups get the most out of DeepSeek. Individuals who want a capable general assistant get free access to DeepSeek through the web and the mobile app, with no API key or billing setup. High-volume developers benefit even more, because RAG pipelines and agent loops replay long shared prefixes on every call, which is exactly the pattern caching rewards.
It is not the right pick for organisations with data residency obligations. The DeepSeek privacy policy states directly that the company collects, processes and stores personal data in the People's Republic of China, which will rule it out for many regulated European and North American teams no matter how the models score. If that describes you, choose a vendor whose hosting jurisdiction already matches your compliance posture.
Strengths
- Genuinely free consumer chat with no message caps or paywall — rare among frontier-class assistants
- Lowest API pricing in the category by a wide margin, with cache-hit input effectively free
- Open MIT-licensed weights allow full self-hosting for data-sensitive workflows
- Strong reasoning and STEM/coding performance that rivals far pricier closed models
Watch out for
- Hosted service processes data on servers in China — a non-starter for many enterprises and privacy-conscious users
- Answer-engine and citation discipline lags Perplexity; no polished dedicated deep-research agent
- No enterprise tier, SLA, or dedicated support on the hosted product
- Web chat can slow during traffic surges, with no guaranteed uptime
Key features
| Feature | What it does |
|---|
| DeepSeek V4 dual-mode | Instant (Flash) and Expert (Pro / Thinking) modes with a 1M-token context window for long documents and codebases |
| Integrated web search | Pulls live web results into answers, letting it act as an answer engine for current-events and research questions |
| Open MIT-licensed weights | Full V4-Flash and V4-Pro weights are downloadable and self-hostable, unmatched by closed rivals |
| Ultra-low-cost API | V4 Flash at $0.14/$0.28 per 1M tokens runs roughly 35–100x cheaper than comparable OpenAI or Anthropic models |
How DeepSeek compares
| Alternative | How it differs |
|---|
| Perplexity | Purpose-built answer engine with far stronger citation discipline and a fast deep-research agent, but Pro costs $20/mo |
| ChatGPT | Broader ecosystem and the longest, most structured deep-research reports, but Plus is $20/mo and it is closed-source |
| Claude | Best-in-class long-document analysis and writing with strong research output, but pricier and no free frontier tier |
Frequently asked questions
Is DeepSeek free?
Yes — the web chat at chat.deepseek.com and the iOS/Android apps are free with no subscription and no documented message cap. Only the developer API is paid, on a pay-as-you-go basis, and new accounts get 5M free tokens.
Where is my data processed?
DeepSeek's hosted chat and API run on servers in China, which raises data-residency and privacy concerns for many enterprises. Self-hosting the open MIT-licensed weights avoids this entirely.
DeepSeek ratings on other platforms
Independent user ratings from third-party review sites, linked here for transparency.
These are not our editorial score, are captured on the date shown, and may have changed since.