DeepSeek pricing depends on how you access the service. DeepSeek advertises its consumer assistant as free, while developer API usage has separate charges. For an API budget, count cached input, uncached input, generated output, and any billable Search or Fetch requests.
Prices reviewed: November 5, 2026. All dollar amounts below are in USD and refer to the direct DeepSeek API.
Independent guide: en.chat-deep.ai is not affiliated with or endorsed by DeepSeek.
DeepSeek API Pricing at a Glance
These are the published rates per 1 million tokens in DeepSeek’s official pricing table.
| API model | Period | Uncached input | Cached input | Output |
|---|---|---|---|---|
deepseek-flash | Off-peak | $0.15 | $0.003 | $0.60 |
deepseek-flash | Peak | $0.30 | $0.006 | $1.20 |
deepseek-v4-pro | Off-peak | $0.66 | $0.022 | $1.98 |
deepseek-v4-pro | Peak | $1.32 | $0.044 | $3.96 |
The Flash identifier serves the V4.1 Flash family. DeepSeek’s Change Log records the October 17, 2026 update to V4.1-Flash-1017 without a price change, and confirms that V4 Pro API service continues after September 14, 2026.
A low cached-input rate applies only to qualifying input tokens. To estimate a complete task, include the remaining input and the generated output, then add any separately billed tools.
Jump to a section
- Input, caching, and output charges
- Peak and off-peak hours
- Search and Fetch pricing
- Cost formula and worked examples
- Billing and balances
- Reducing costs and other access options
- Pricing FAQ
What Does DeepSeek Charge For?
A token is a unit the model processes, rather than a fixed number of words. Text, code, and images can produce different token counts. DeepSeek’s token usage guide says the usage returned by the API is the reference for the actual amount processed. Word-count conversions are only estimates.
Input includes the context you send
Input can contain system instructions, the user’s question, supplied documents, and conversation history. DeepSeek’s multi-round conversation guide explains that chat-completion requests must include the history needed for the next turn.
This matters for budgeting: a short follow-up question can arrive with a much longer context. Count what your application sends on each request, rather than counting only the words a person types into the chat box.
Cached input is part of total input
DeepSeek’s context caching documentation describes automatic caching of reusable input prefixes. A hit requires a complete match to a prefix unit already persisted in the cache. Similar wording or repeated subject matter does not guarantee a hit, and caching operates on a best-effort basis.
For example, if a request contains 2,000 input tokens and 1,800 are cache hits, the remaining 200 are cache misses. Apply the cached rate to 1,800 and the uncached rate to 200. Charging all 2,000 at the uncached rate and adding 1,800 cached tokens would count part of the input twice.
Output includes reasoning usage
The Chat Completions reference reports generated usage in completion_tokens, with reasoning tokens shown as a breakdown within completion details. Use the completion total once when calculating output cost; do not add the reasoning breakdown again.
DeepSeek’s thinking-mode guide documents reasoning controls and notes that thinking is enabled by default. A brief visible answer can therefore have more generated tokens than its final text suggests.
Peak and Off-Peak Pricing Explained
DeepSeek publishes two peak windows: 01:00–04:00 UTC and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other times are off-peak, including weekends and Chinese public holidays. The listed off-peak model rates are half the peak rates.
Use the official UTC schedule when planning flexible work. For a US team, local dates can differ from UTC dates, and daylight saving time changes local conversions. Schedule against UTC and the applicable holiday rules instead of assuming a fixed local evening always qualifies.
For a workload spread across both periods, calculate each period separately. If you do not know when requests will run, a peak-rate estimate provides a conservative model-cost scenario for the same assumed token volume.
DeepSeek Search and Fetch API Pricing
The official Search and Fetch guide lists separate request-based charges:
| Service | Billing unit | Price |
|---|---|---|
| Search API | 1,000 billable requests | $0.75 |
| Fetch API | 1,000 billable requests | $0.75 |
A billable request returns HTTP 200 with usable results: a nonempty result list for Search, or page content for Fetch. The guide excludes failed requests, empty search results, and timeouts from these tool charges. The number of Search results requested does not change the per-request price.
These charges also apply to successful server-side Search and Fetch calls through supported API integrations, including DeepSeek Harness. The guide publishes one tool rate, without a separate off-peak discount.
If retrieved material is then supplied to a model, include that material in your model-input estimate too. Your budget therefore needs both the retrieval requests and the model work that uses their results.
How to Calculate Your DeepSeek API Cost
For one model and one pricing period, use the three token rates from the first table:
Model cost = (cached input tokens × cached-input rate + uncached input tokens × uncached-input rate + completion tokens × output rate) ÷ 1,000,000 Search and Fetch cost = (billable Search requests + billable Fetch requests) × $0.75 ÷ 1,000 Estimated API cost = model cost + Search and Fetch cost
Repeat the model calculation for each model and time period, then add the results. When starting with total input tokens, subtract cached input to obtain uncached input.
| Chat Completions usage field | How to use it |
|---|---|
prompt_cache_hit_tokens | Multiply by the cached-input rate. |
prompt_cache_miss_tokens | Multiply by the uncached-input rate. |
completion_tokens | Multiply by the output rate, including reasoning usage already in this total. |
total_tokens | The combined input and output count; it cannot be priced accurately using one rate. |
These field relationships are documented in the API reference. Other API formats can use different field names, so use the usage schema for the endpoint you call.
DeepSeek API Cost Examples
The following are hypothetical calculations using the November 5, 2026 rate table. They are planning examples, not measured workloads or promised bills. Output counts mean total billable completion tokens. Taxes, payment-provider charges, account credits, and your application’s hosting costs are outside these estimates.
Example 1: One short Flash request
Assume a Flash request uses 2,000 uncached input tokens and 500 output tokens during off-peak hours, with no Search or Fetch calls:
- Input: 2,000 ÷ 1,000,000 × $0.15 = $0.00030.
- Output: 500 ÷ 1,000,000 × $0.60 = $0.00030.
- Total model cost: $0.00060 per request.
Preserve these small decimal amounts while estimating volume. Rounding every request to cents before adding them would distort the total.
Example 2: A monthly workload of 10,000 requests
Keep the same 2,000 uncached input tokens and 500 output tokens per request. Across 10,000 requests, that is 20 million input tokens and 5 million output tokens, with no cache hits or billable tools.
| Model and timing | Input cost | Output cost | Monthly model estimate |
|---|---|---|---|
| Flash, all off-peak | $3.00 | $3.00 | $6.00 |
| Flash, all peak | $6.00 | $6.00 | $12.00 |
| V4 Pro, all off-peak | $13.20 | $9.90 | $23.10 |
| V4 Pro, all peak | $26.40 | $19.80 | $46.20 |
This compares identical assumed token volumes. Actual models can produce different output lengths or need different numbers of attempts to complete a task. Evaluate cost alongside the quality of the completed work.
If 6,000 of these identical Flash requests run off-peak and 4,000 run during peak hours, the model estimate is $3.60 + $4.80 = $8.40.
Example 3: The same workload with cached input
Suppose 90% of the 20 million input tokens in Example 2 qualify as cache hits. Keep Flash, off-peak timing, and 5 million output tokens:
- Cached input: 18,000,000 ÷ 1,000,000 × $0.003 = $0.054.
- Uncached input: 2,000,000 ÷ 1,000,000 × $0.15 = $0.30.
- Output: 5,000,000 ÷ 1,000,000 × $0.60 = $3.00.
- Total: $3.354, or approximately $3.35.
That is a 44.1% reduction from the $6.00 uncached scenario. Output still costs $3.00, which explains why a high input-cache percentage does not produce an equally large reduction in the whole bill. The assumed 90% cache-hit share is an example, not a guaranteed result.
Example 4: Adding Search and Fetch
Return to the $6.00 uncached, off-peak Flash workload in Example 2. Add 10,000 billable Search requests and 20,000 billable Fetch requests during the month:
- Search: 10,000 ÷ 1,000 × $0.75 = $7.50.
- Fetch: 20,000 ÷ 1,000 × $0.75 = $15.00.
- Combined model and tool estimate: $6.00 + $7.50 + $15.00 = $28.50.
This assumes the model-token totals already include all retrieved content submitted to the model. If research adds more input, generated output, or model calls, add that usage as well.
How DeepSeek API Billing and Balances Work
The Open Platform Terms describe paid usage consuming account balance and allow service suspension when the balance is insufficient. Payment options and any free allowance depend on the information displayed by the platform.
DeepSeek’s pricing rules say granted balance is used before topped-up balance when both are available. The balance documentation distinguishes these amounts and identifies the balance currency. A documented granted-balance field does not establish that every account receives free credits.
For budgeting, distinguish the money added to an account from the usage consumed during a month. A $10 top-up, for example, is an account-funding amount; it is not a fixed number of requests or a monthly service package.
Refund requests are subject to review under the platform terms. If approved, they cover the remaining unspent amount after necessary costs, with the remaining balance refunded together. The terms exclude partial refunds and refunds of spent amounts. Review the payment and refund terms before funding an account.
How to Reduce DeepSeek API Costs
Start with the part of your workload responsible for the largest charge. The most useful change for a long-document workflow may differ from the most useful change for an assistant that writes lengthy answers.
- Match the model to acceptable results. Compare the cost of completing representative tasks, including any additional attempts or review time.
- Remove irrelevant input. Send the context needed for the task and manage growing conversation histories deliberately.
- Keep reusable prefixes stable. Where the task permits, place shared instructions and context consistently. Use reported cache hits to assess the benefit.
- Request an appropriate output length. A concise answer or structured result can reduce unnecessary generation. Check that the chosen output limit still allows the task to finish.
- Choose reasoning settings deliberately. Review the supported controls in the thinking-mode guide and measure how they affect both usefulness and completion usage.
- Schedule flexible work off-peak. Apply the published UTC windows and holiday exceptions when the task can wait.
- Track tools as well as tokens. Repeated searches or fetches can become a larger budget item than the model response itself.
Keep separate totals for each model, rate period, token category, and billable tool. That makes it easier to explain a cost increase and assess whether a change actually reduced spending.
Free Chat, Direct API, and Third-Party Pricing
The free consumer assistant is intended for interactive use through DeepSeek’s own interfaces. It does not turn the developer API into a free service. For software that sends automated requests, estimate the API usage described above.
If another application or hosting provider offers a DeepSeek model, use that provider’s rate card and billing terms. A subscription can include application features or a usage allowance that differs from direct API billing. Confirm who receives your payment and which service is being measured before comparing prices.
Running a downloadable model yourself is a separate cost calculation. Hardware, hosting, electricity, and maintenance can matter even when model files are available without a download fee.
DeepSeek Pricing FAQ
How much does 1 million DeepSeek tokens cost?
There is no single price for a mixed total. Identify the model, pricing period, and how many tokens are uncached input, cached input, or output. Apply the corresponding rates in the table and add those amounts.
Does caching also discount output tokens?
The cache-hit rate applies to qualifying input. DeepSeek’s context caching guide explains that output is still generated. Keep output as a separate part of the estimate even when most input is cached.
Can I estimate the cost without an API key?
Yes. The formula and worked examples can be used with assumed token and tool volumes without making an API request. If you later have usage records, replace the assumptions with the reported counts to improve the estimate.
Are all failed API requests free?
The Search and Fetch guide explicitly excludes failed, empty-result, and timed-out tool requests from its charges. That rule should not be generalized to every model request or interrupted generation. Check the usage and billing records for the service involved.
Why do other guides show different DeepSeek prices?
A guide may describe an older model, a different time period, a promotion, or another provider. R1- or V3-era tables should not be applied automatically to the model IDs above. Match the model identifier and billing date, and consult the official Change Log for changes.
For your own budget, start with expected token volumes, choose the applicable rate periods, and add billable tools. Recheck the official DeepSeek pricing page before committing to a spending plan.