What is DeepSeek?
DeepSeek is a Chinese AI lab whose open-weight models drew global attention by reaching near-frontier quality at dramatically lower training and inference cost. Its general model (DeepSeek-V3) and reasoning model (DeepSeek-R1) are both downloadable and very cheap to call by API.
You can use DeepSeek through its web chat, its low-cost API, or by self-hosting the open weights. R1, in particular, is a strong, transparent reasoning model that shows its chain of thought and competes with far more expensive offerings.
DeepSeek is the model to evaluate when cost is the dominant constraint and you want open weights you can host, provided you've thought through the data-governance questions that come with a China-based hosted service.
What it's best for
- Cost-sensitive workloads: its API is among the cheapest of any capable model, and the open weights cost only your own compute.
- Reasoning tasks: DeepSeek-R1 shows its work and performs strongly on math, logic, and coding problems.
- Self-hosting: open weights mean you can run it entirely in your own environment.
- High-volume processing where price-per-token decides the architecture.
- Research and experimentation on a capable model without a large budget.
Where it falls short
- Sensitive or regulated data on the hosted service. DeepSeek's hosted API and app run in China, with the data-residency and governance implications that follow. Self-hosting avoids this.
- Topics subject to Chinese content restrictions. The hosted model declines or steers certain political subjects.
- Teams needing a polished consumer feature set and mature enterprise support.
- Guaranteed uptime and SLAs at the level of the big US providers (on the hosted API).
Three ways in
Use the web app at chat.deepseek.com for a free assistant. For building, call the DeepSeek API (OpenAI-compatible, so most SDKs work with a base-URL change). For full control, download the open weights and self-host.
Choose V3 for general chat and tasks, R1 when you want explicit step-by-step reasoning.
Self-hosting for data control
The open weights are published on hubs like Hugging Face and run under common runtimes. Self-hosting is the recommended path for any North-American team with data-sensitivity requirements, because it keeps inference off China-based infrastructure.
Distilled smaller versions of R1 exist that run on more modest hardware while keeping much of the reasoning ability.
Getting better answers
With R1, let it reason and don't force a one-line answer. The visible chain of thought is useful for checking how it reached a conclusion.
Because the API is OpenAI-compatible, you can A/B it against other models in existing code by swapping the base URL and key.
What DeepSeek costs
Approximate, in USD, as of January 2026. Prices change often. Confirm on the official site before you rely on them.
Open weights
$0 (self-host)
Download and run V3 or R1 yourself; you pay only for your own compute.
Web app
$0
Free assistant at chat.deepseek.com, subject to limits.
API
Very low, usage-based
Among the cheapest per-token pricing of any capable model; off-peak discounts.
Example prompts
Copy these into DeepSeek as starting points, then adapt them to your task.
Show the reasoning
Solve the problem below with DeepSeek-R1. Show your full reasoning, then give a final answer on its own line. If multiple interpretations are possible, address each.
Cheap high-volume extraction
Extract {fields} from each record below as JSON. Be strict: return only valid JSON, one object per record, null for missing values.
Swap-in test
Rewrite this OpenAI API call to use the DeepSeek API instead (base URL and model name), keeping everything else the same.
Governance check
We're considering DeepSeek for an internal tool that touches customer data. List the data-governance questions we must answer before using the hosted API, and what changes if we self-host the open weights instead.
DeepSeek
common questions.
Direct answers to the questions we get asked the most. If yours isn't covered, write to the team.