DeepSeek Pricing 2026

DeepSeek pricing is two unrelated answers

Almost every disagreement about what DeepSeek costs comes from two groups of people answering different questions with the same word. For someone who wants to use DeepSeek, the price is $0 and there is no tier to buy. For someone calling the API, it is metered per token through a developer account, and the number depends on the model, on prompt caching, and now on the time of day.

The middle row is deliberately blank, and the next section explains why rather than hiding it.

Why this page quotes no DeepSeek token rate

DeepSeek's API documentation refuses automated requests from this network, so no figure on its rate card could be read directly on 2026-08-19. That is the same reason our verification standards page already gives for publishing no DeepSeek token prices, and this page keeps to it.

Three separate sources were available and they do not agree. This site's own quarterly specification table carries one pair of figures. A researched draft written earlier this month carries a different pair, against model names the live documentation no longer uses. A read-through fetch of the vendor's own documentation returned a third answer, under different model identifiers again, and showed the rate card had been restructured on 2026-08-16 into separate peak and off-peak prices. Three answers, no direct read, and a card that moved inside the last week.

Publishing one of the three would look more helpful and would be a coin toss. A price you cannot verify is a price that propagates by copying, and copying is how a stale figure outlives the thing it described. So the structure is reported here and the numbers are not:

• There are two current model tiers, a cheaper and faster one and a more capable one, and the gap between them is roughly threefold rather than marginal.
• Billing now varies by time of day. An off-peak window is priced materially below the peak rate for the same tokens, which is unusual in this market and is worth designing batch work around.
• Cache hits cost a small fraction of cache misses, on the order of a few percent rather than a few tens of percent. This dominates everything else and is covered below.

Read the current figures off DeepSeek's own API documentation before you budget against them. For the vendors whose pages did answer, the plan tables on Claude's pricing page and Google's AI plan pricing carry a source and a date next to every number.

The free app: limited by scope, not by volume

The consumer product is what most of this search is actually about, and it is unusually simple. There is no Plus plan, no Pro plan and no message quota to buy your way past, because there is nothing to sell. The practical ceiling is fair-use throttling when the service is busy, which arrives as slower responses rather than a hard stop.

That is a genuinely different posture from the rest of the market. The free tiers at OpenAI, Anthropic and Google exist to sell the tier above them: the free model is a smaller one, the good model is metered, and the friction is the product. DeepSeek's consumer app is not a funnel, because there is nothing at the end of it.

The limitation is scope. A free DeepSeek account gives you one vendor's model family. It does not give you a different model for long-form review, or for large multimodal inputs, or for live web context, and nothing carries what it knows about your work from one of those to another. For a lot of people that is fine and the free app is the right answer. For anyone whose work spans model strengths, it is a component rather than a replacement.

Two specifications that change what the model is for

Two capability figures matter more than the rate card here, and unlike the prices they are stated plainly in the vendor's own documentation: a 1 million token context window and a 384,000 token maximum output.

The output ceiling is the unusual one. Competing APIs commonly cap a single response somewhere between 8,000 and 64,000 tokens, which forces anything book-length to be chunked, generated in pieces and stitched back together, with all the continuity problems that creates. Being able to emit a very long document in one call removes a real engineering problem rather than improving a number. If that is your workload, it is the reason to look at this vendor, and it survives whatever the rate card does next.

For how that context figure reads against the rest of the field, context window limits across the major models puts the published numbers side by side.

Prompt caching is the decision, not the rate

The most consequential number on DeepSeek's pricing page is not the headline rate, it is the cache-hit rate, which bills at a small fraction of a cache miss. That matters because most production workloads are prefix-heavy by construction. A support agent sends the same instructions, tool definitions and reference material on every call, and only the user's message changes. A document assistant re-sends the same document throughout a conversation. In those shapes the hit rate can sit above 90%, and the effective input cost collapses toward the cached rate.

The guidance that follows is boring and worth more than any rate comparison: put your stable content first and your variable content last. A prompt assembled with the user's question at the top and the system context underneath cannot cache its prefix, and pays the full miss rate on every call for no benefit at all. This is the most common source of a DeepSeek bill that is larger than expected, and it is a prompt-ordering bug rather than a pricing problem.

Where hosting matters more than price

DeepSeek's hosted API and consumer app process data on Chinese infrastructure. For regulated industries, for work covered by client confidentiality, and for organisations with a policy on where work product is processed, that is a documented consideration with a factual answer for your situation rather than a verdict on the model. Self-hosting the open weights removes it completely, since nothing leaves your own environment; using the hosted service does not.

The open weights are the part that makes this vendor structurally different from a closed lab: a team with GPU capacity can run the model itself, and none of the hosted pricing applies on that path. What running an LLM on your own hardware involves works through what that actually costs in practice.

DeepSeek is rarely the expensive line

The reason a DeepSeek pricing page ends up being about something else is that DeepSeek is not usually the variable in anyone's AI spend. The consumer app is free. A heavy individual's API usage, at any of the three rate cards above, lands in the low single-digit dollars a month.

The variable is the general assistant subscription sitting next to it for the work DeepSeek is not the right model for, and possibly a second one after that. That is the line worth re-pricing. Perspective AI is $14.99/mo for models from OpenAI, Anthropic, Google, xAI, DeepSeek and Mistral in one app, with mid-conversation switching and memory that carries across them, which is below the price of the single frontier plan most DeepSeek users are still paying for. It keeps DeepSeek available for the work it is best at, and it makes the model choice a per-task decision rather than one made once at signup.

The honest limit on that argument: if your DeepSeek use is the hosted API at volume, a flat consumer subscription is not the same product and does not replace it. This page's consolidation case is about the subscription beside it, not about the API.

The short version

Consumers pay nothing and there is nothing to upgrade to. Developers pay per token on a card this page will not quote, because the vendor's documentation refuses automated requests, three available sources disagree, and the rates were restructured into peak and off-peak bands on 2026-08-16. What is verifiable and stable is the shape: two model tiers, a time-of-day split, cache hits at a small fraction of misses, a 1 million token context window and a 384,000 token output ceiling. Design for the cache, read the current rates off the vendor before you budget, and price your whole setup rather than one model at a time.