How DeepSeek API pricing works
There's no subscription and no monthly fee. You add credit to your account on the API platform, and each request deducts its cost from that balance. The cost of a request depends on three things:
Input tokens
Your prompt, system message and any history. Repeated input can be billed at a lower cached rate.
Output tokens
The reply, plus R1's reasoning. Output costs more per token than input.
What is a token?
A token is a small piece of text: a short word, part of a longer word, or a punctuation mark. Models read and write in tokens, so that's what you're billed for. These rough rules of thumb help you estimate:
| Text | Approximate tokens |
|---|---|
| 1 English word | about 1.3 tokens |
| 750 English words | about 1,000 tokens |
| A short chat message | 20–60 tokens |
| One page of a document | 500–800 tokens |
| A typical email | 150–400 tokens |
| Code | More tokens per line than prose, because of symbols and spacing |
| Chinese, Sinhala, Tamil and other non-Latin scripts | Varies. Often more tokens per character than English |
Price per model
Rates per 1 million tokens, in US dollars. Models are listed from cheapest to most expensive.
| Model | Input · cache hit | Input · cache miss | Output | Best for |
|---|---|---|---|---|
| V4.1 FlashCheapest | Default for most work | |||
| CoderCode | Completion - code generation | |||
| R1Reasoning | Maths, logic, hard bugs | |||
| V4 ProFlagship | Long, high-stakes work |
Cache hits and misses: the biggest saving
DeepSeek automatically caches the beginning of your requests. If a new request starts with exactly the same text as an earlier one, that shared part is billed at the much lower cache hit rate. Only the new text pays the full cache miss rate.
How to get more cache hits
- Put fixed content (system prompt, instructions, examples, reference documents) first.
- Put the changing content (the user's question) last.
- Keep the fixed part word-for-word identical. Even a timestamp at the top breaks the match.
- In conversations, earlier turns stay the same, so long chats get cache hits automatically.
The response's usage field shows how many input tokens were cache hits and how many were misses, so you can check it's working.
Reasoning, search and other costs
| What | How it's billed | What to watch |
|---|---|---|
| R1 reasoning | The chain of thought counts as output tokens, even though it isn't the final answer. | A short answer can still use thousands of thinking tokens. Send only hard tasks to R1. |
| Web search - fetch | The search results and page text added to the request count as input tokens. | Use them only when you need current information. See search - fetch. |
| Conversation history | Every earlier message you send again is billed as input. Unchanged history usually counts as cache hits. | Summarise very long chats. |
| Tool calls | Tool descriptions and tool results are part of the input. | Keep tool descriptions short. |
| Failed requests | Requests rejected with an error (for example 401 or 402) aren't charged. | Requests cut short by max_tokens are charged for the tokens used. |
DeepSeek API cost calculator
Pick a model and describe your usage. The estimate updates as you type. If the rates are blank, copy the current numbers from the API platform into the three rate boxes.
For R1, include the reasoning tokens in "output tokens per request". They're often several times longer than the answer.
Worked examples
Typical token counts for common apps. Click Load into calculator to see the cost at current rates.
| Use case | Model | Requests / day | Input / request | Output / request | Cache hit | |
|---|---|---|---|---|---|---|
| Customer support bot | V4.1 Flash | 2,000 | 1,200 | 250 | 70% | |
| Document summariser | V4.1 Flash | 300 | 6,000 | 400 | 10% | |
| Email classifier / extractor | V4.1 Flash | 10,000 | 400 | 40 | 80% | |
| IDE autocomplete (one developer) | Coder | 1,500 | 1,500 | 60 | 60% | |
| Maths tutor | R1 | 200 | 300 | 2,500 | 50% | |
| Long report writer | V4 Pro | 20 | 8,000 | 3,000 | 20% |
Billing, top-ups and your balance
| Question | Answer |
|---|---|
| How do I pay? | Top up a prepaid balance under Billing on the API platform. Each request's cost is deducted from it. |
| What happens at zero balance? | Requests fail with error 402 Insufficient balance until you top up. See API errors. |
| Is there a free trial? | Promotional credit is offered from time to time. Check your console. The chat app is always free. |
| Where do I see my spending? | The Usage page in the console breaks spending down by day and model. Each API response also includes token counts. |
| Can I set a spending limit? | Because the API is prepaid, your balance is the limit. Top up in small amounts while you test. |
| Is the API account the same as my chat account? | No, they're separate. Having trouble signing in? See the login guide. |
How to lower your DeepSeek API bill
| Technique | Typical saving | How to do it |
|---|---|---|
| Default to Flash | Large | Send everything to V4.1 Flash first. Send only flagged hard cases to R1 or V4 Pro. |
| Maximise cache hits | Large on input | Put the fixed prompt first, keep it identical, and put the variable text last. |
| Cap output length | Medium | Set max_tokens, and ask for "under 100 words" or "JSON only". |
| Trim context | Medium | Send only the relevant part of a document, and summarise old conversation turns. |
| Avoid retries | Small to medium | Use JSON mode and clear formats so replies parse first time. |
| Use search only when needed | Varies | Fetched pages add input tokens, so turn search on only for current-information questions. |
| Self-host at very high volume | Can be large | Run open-weight models on your own hardware. See open weights. |
Free vs paid: the chat app, the API and Harness
| Product | Cost | Who it's for |
|---|---|---|
| Chat app (web, iOS, Android) | Free | Anyone chatting, writing, researching or learning |
| API | Prepaid, per token | Developers building apps, bots and automations |
| Harness Desktop | Free public preview; can use your account or an API key | Hands-on work in your files: coding, documents, spreadsheets, scheduled tasks, plugins |
| Self-hosted open weights | Free model; you pay for hardware | Teams needing privacy or very high volume |
DeepSeek API pricing - frequently asked questions
How much does the DeepSeek API cost?
It depends on the model and on how many tokens you send and receive. Rates are quoted per million tokens, with separate prices for cache-hit input, cache-miss input and output. See the price table and use the calculator.
Which is the cheapest DeepSeek model?
V4.1 Flash, and the 1017 update also lowered its cost per task.
Why is R1 more expensive per request?
Its reasoning is billed as output tokens, and on hard problems it can write thousands of them before it answers.
What is a cache hit?
When the start of your request matches an earlier one exactly, that part is billed at a lower rate. Put fixed instructions first to benefit.
Is the DeepSeek chat app free?
Yes. Pricing applies only to the API. The chat app on the web, iOS and Android is free.
Do I need a credit card for a subscription?
There's no subscription. You top up a prepaid balance and pay only for what you use.
Am I charged for failed requests?
Requests rejected with an error aren't charged. Replies cut short by max_tokens are charged for the tokens used.
Is this DeepSeek's official price list?
No. DeepSeekR1.com is an independent guide, and the official prices are on the API platform. See about us, our disclaimer, privacy policy and terms.