Pricing and plans

DeepSeek's pricing model is unusual among the major labs: there is no paid consumer subscription. The chat app is free, the API is metered per token, and because the weights are open there is a third route — pay somebody else to host the model, or host it yourself. This lesson puts the three side by side.

Route 1: the free app

For an individual doing writing, research, coding help and everyday questions, the free app is the right starting point and often the end point. You get the current chat model, the DeepThink reasoning mode, web search and file upload at $0. The trade-offs are the ones listed in Lesson 1: capacity limits during busy periods, a thinner feature set than ChatGPT or Gemini, and data handled under DeepSeek's own policy.

If you are currently paying for a consumer AI subscription and mainly use it for text and code, the honest test is to spend a week on the free DeepSeek app and see what you miss. Many people find the answer is nothing, or only voice mode and image generation. Our cheapest AI tools page lists what the free tiers of each product actually include.

Route 2: the API

The API is for anything you want to run without a human clicking: scripts, integrations, apps, batch jobs, and IDE tools that need a key. You pay per token at the rates in the Lesson 2 ladder, with two discounts unique to DeepSeek — cache-hit input tokens are billed at a much lower rate, and the official API publishes off-peak windows with reduced prices. The exact figures live on the official pricing page.

A sense of scale, using catalogue prices at the time of writing: a developer sending 200 requests a day to V3.2 at 3,000 input and 800 output tokens each uses 18M input and 4.8M output tokens a month. That is roughly 18 × $0.27 + 4.8 × $0.40 ≈ $6.78 a month before caching. The same load on R1 0528 would be about 18 × $0.50 + 4.8 × $2.15 ≈ $19.32, and realistically more because of reasoning tokens.

Route 3: open weights

Because the weights are open, dozens of hosts serve DeepSeek models. Our price ladder already includes rows from those hosts, which is why you see several prices for what looks like the same model. Third-party hosts are worth it when you need a specific region for data reasons, better uptime than the official API during spikes, or a dated checkpoint pinned for reproducibility.

Self-hosting only pays off at volume. The full-size V3 and V4 models need multiple high-memory GPUs, so a rented cluster costs more per month than most teams would spend on the API. The distilled R1 variants and smaller checkpoints are a different story — they run on a single workstation GPU and are genuinely free after the hardware. If you are already paying for idle GPUs, that is the moment to consider it.