DeepSeek review

What it is

DeepSeek is a Chinese AI lab whose open-weight reasoning models reset expectations about what frontier-adjacent capability should cost. There is a free consumer chat app, a very cheap API, and downloadable weights you can run on your own hardware.

Where it shines

Price per unit of capability is the headline, and it is not close. For teams building products where inference cost determines the unit economics — summarisation at scale, classification pipelines, in-app assistants — routing the bulk of traffic to DeepSeek and reserving an expensive model for hard cases is a standard cost-control pattern in 2026.

The reasoning models are genuinely strong at maths, logic and code. If your workload is structured problem solving rather than polished prose, the gap to the most expensive options is small enough that most users would not notice in blind testing.

Open weights are the third advantage. Regulated organisations that cannot send data to any external API can run the model inside their own network, which is simply not an option with the closed frontier labs.

Where it falls short

The hosted service raises legitimate governance questions. Data sent to the consumer app and hosted API is processed on infrastructure subject to Chinese jurisdiction, and a number of governments and enterprises have restricted its use on official devices. If you work somewhere with a data policy, read it before signing up — and consider self-hosting instead.

English prose is the other weakness. It is clear and correct but flat, and it needs more editing than Claude or ChatGPT output. Product features are minimal: there is no meaningful agent framework, no file analysis suite, no voice mode. You are buying a model, not a product.

Pricing in plain English

Consumer chat is free. API pricing starts at a small fraction of frontier rates per million tokens, with further off-peak discounts and context caching. Self-hosting costs only your own compute.

Who should use it

Developers optimising inference cost, teams that need on-premise deployment, and anyone doing maths-heavy or code-heavy work on a budget. Avoid the hosted service for confidential business data unless your policy explicitly permits it.