DeepSeek pricing depends on how you access the service. The official mobile assistant is advertised as free to use, while the DeepSeek API charges for token consumption. Your API bill depends on the model, input and output volume, cache hits, and whether usage falls within peak or off-peak hours. Web-retrieval workflows can also incur separate Search and Fetch request fees.
The official consumer pages reviewed do not advertise a separate monthly or annual paid subscription. In the API tables below, “Pro” identifies a model—not a monthly subscription tier. A free consumer app and a funded developer account are different things.
Token prices and billing documentation checked: October 2, 2026. Search and Fetch fees and limits rechecked: October 19, 2026. All monetary amounts below are in US dollars.
Current DeepSeek API prices per million tokens
These are first-party API rates, not prices charged by another app or hosting provider. Each amount applies to 1 million tokens in the specified category, not to 1 million messages or requests.
DeepSeek V4.1 Flash pricing
Use the API identifier deepseek-flash. The following rates are published in DeepSeek’s V4.1 Flash pricing announcement, effective September 10, 2026.
| Token category | Off-peak | Peak |
|---|---|---|
| Input: cache miss | $0.15 | $0.30 |
| Input: cache hit | $0.003 | $0.006 |
| Output | $0.60 | $1.20 |
USD per 1 million tokens. The cached-input price applies only to input tokens recorded as cache hits.
DeepSeek V4 Pro pricing
Use the API identifier deepseek-v4-pro. These rates were introduced in the V4 Pro pricing update, effective August 16, 2026.
| Token category | Off-peak | Peak |
|---|---|---|
| Input: cache miss | $0.66 | $1.32 |
| Input: cache hit | $0.022 | $0.044 |
| Output | $1.98 | $3.96 |
USD per 1 million tokens for the official DeepSeek V4 Pro API.
V4 Pro remains available. Although an earlier announcement proposed a September transition away from Pro, the updated official change log confirms continued API service after September 14, 2026, with its billing method unchanged.
The same change log explains that the older Flash identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to DeepSeek-V4.1-Flash-1017. Using a legacy request name does not preserve the retired model or its old pricing.
When do peak and off-peak prices apply?
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and those holidays. The holiday exception is specified in the official pricing footnote.
Off-peak rates are half the corresponding peak rates. For work that can wait, such as a scheduled document-processing job, timing can therefore change the token bill without changing the model. For a workload spread across both periods, calculate the usage in each period separately rather than applying the lower rate to the entire month.
How to calculate your DeepSeek API cost
Separate uncached input, cached input and billed output. Multiply each token count by its applicable price, add the amounts, and divide by 1 million:
Cost in USD = (uncached input tokens × uncached input rate + cached input tokens × cached input rate + billed output tokens × output rate) ÷ 1,000,000
This formula uses actual token counts and rates expressed per million tokens. Apply it separately for each model and pricing period, then add the results. Cached and uncached input must divide the input total between them—not duplicate it. The formula covers token charges; add any billable Search and Fetch request fees separately.
Example: 10,000 API calls in a month
Consider a hypothetical 30-day workload of 10,000 calls, each containing 2,000 uncached input tokens and generating 500 billed output tokens. That gives 20 million input tokens and 5 million output tokens for the month. The output assumption includes any reasoning tokens; it is not merely the length of the visible final answer.
Using the verified rates above, Flash’s off-peak calculation is (20 × $0.15) + (5 × $0.60) = $6.00. Here, the quantities are already expressed in millions, so no further division is needed.
| Model | All calls off-peak | All calls at peak rates |
|---|---|---|
| V4.1 Flash | $6.00 | $12.00 |
| V4 Pro | $23.10 | $46.20 |
Calculated examples for 20 million uncached input tokens and 5 million output tokens. These are not subscription plans or typical customer bills.
For Flash off-peak, one call with these assumptions costs $0.0006. A small nonzero charge is still a charge; rounding every request to the nearest cent would incorrectly make it appear free.
These estimates cover the specified model usage only. They exclude separate hosting, external tools, provider fees and any applicable taxes. Additional model calls or longer responses must be included in a real budget.
Example: the same workload with cached input
Now keep the same Flash workload and off-peak schedule, but assume its actual usage records show that 80% of input tokens were cache hits: 16 million cached input tokens, 4 million uncached input tokens and the same 5 million output tokens.
The calculation becomes (16 × $0.003) + (4 × $0.15) + (5 × $0.60) = $3.648, or approximately $3.65.
That is a 39.2% reduction from the uncached $6.00 example. It is not a 98% reduction of the whole bill: the output charge remains unchanged. The assumed cache-hit share covers the complete workload, including its initial cache misses; it is an illustration, not a promised hit rate.
What counts toward the bill?
Search and Fetch request fees. DeepSeek’s Search & Fetch documentation, checked October 19, 2026, lists separate charges for POST /v1/tools/search and POST /v1/tools/fetch. Add these to any model-token charges incurred by the workflow.
| API | Billable unit | Price (USD) |
|---|---|---|
| Search | 1,000 successful requests with non-empty results | $0.75 |
| Fetch | 1,000 successful requests returning page content | $0.75 |
A request is billable only when it returns HTTP 200 and produces a result. Failed requests, empty search results, and timeouts are not billed. Search’s count changes how many results are returned, not the per-request price. For example, 1,000 billable Search calls plus 1,000 billable Fetch calls cost $1.50 in request fees, before any model-token charges.
The account-level limit is 100 requests per second for Search and 25 for Fetch, independently, across the account’s API keys. These request-rate limits are separate from model concurrency; Fetch also has an additional per-domain limit.
Search and Fetch request-fee notice: Server-side tool calls that use Search and Fetch through DeepSeek Harness, Claude Code, or Codex are currently free of request charges. Per-request billing starts at 00:00 Beijing time on November 1, 2026. This notice does not waive model-token charges or establish free direct standalone Search or Fetch usage.
Tokens are not messages or words
A token is a unit the model processes, not a fixed number of words or a complete message. Text-to-token estimates vary with the content and model. DeepSeek’s token-usage documentation advises using the consumption reported by the API for the actual count.
For Chat Completions, the API response reference exposes prompt_cache_miss_tokens, prompt_cache_hit_tokens and completion_tokens. These correspond to the three categories in the calculation. Do not add prompt_tokens again: it already combines the two input categories.
Likewise, total_tokens mixes input and output, which have different prices. Multiplying that total by one advertised rate is not an accurate cost calculation.
Caching discounts qualifying input, not the entire request
DeepSeek’s context caching is enabled by default. However, a cache hit requires a matching prefix that has already been stored. Merely asking about the same topic, or sending similar wording, does not guarantee a discount.
A practical approach is to keep reusable instructions and reference material stable near the beginning of related requests, with changing questions afterward. Then inspect the reported hit counts rather than assuming the intended reuse succeeded. DeepSeek describes caching as best-effort, not a guaranteed percentage.
The discounted price applies to qualifying input tokens. The model still generates output, and that output has its own charge.
Reasoning and multi-step tasks can increase consumption
In the Chat Completions usage schema, reasoning tokens are a breakdown within completion usage. Use the total billed completion count rather than charging only for the visible answer—or adding the reasoning count a second time.
Thinking is enabled by default in the current API documentation. DeepSeek’s thinking-mode guide documents effort controls and workflows involving multiple model calls. A single user task may therefore generate several billable calls before producing its final result.
For budgeting, measure the full task, including its intermediate calls and supplied context. Test whether a lower thinking effort or a shorter requested response still meets your quality requirements. Reducing output indiscriminately can produce incomplete answers and additional follow-up calls.
For image workloads, avoid treating one image as a fixed number of text tokens. DeepSeek’s image-token guidance relates consumption to image processing and dimensions; the returned usage remains the reference for your calculation.
API payments, free credit and refunds
Is there a fixed monthly API fee?
The published API model is usage-based, not a fixed monthly message allowance. Under the Open Platform payment terms, paid usage consumes your account balance, and insufficient funds may lead to service suspension. Adding credit funds usage; it does not purchase an unlimited monthly plan.
Before paying, review the top-up amount, payment options and conditions displayed in your own account. Do not treat a suggested top-up amount as a mandatory minimum, or assume that terms attached to promotional credit also apply to purchased credit.
Does every account receive free API tokens?
A universal free API allowance could not be confirmed in the official pages reviewed. The billing rules recognize both topped-up and granted balances and say that granted funds are consumed first when both exist. That is not a promise that every new account receives a particular amount.
Only include a free allowance in your budget after confirming that it is available to your account, which services it covers, and whether an expiry date applies. Developers can also monitor remaining funds through the documented Get User Balance endpoint.
Can unused credit be refunded?
DeepSeek’s published refund policy requires a request and review. If approved, the remaining unspent amount is refunded after necessary costs such as handling fees. The policy does not support partial refunds, and already-consumed amounts are not refundable under that policy. Approval is not automatic.
How to keep spending predictable
Start with a representative sample of your actual workload—not an assumed “average user.” Record the model, input categories, output consumption, pricing period and number of calls needed for a successful result. Use those measurements to project a normal month and a heavier-usage scenario.
Set application-side request budgets, output limits and alerts appropriate to your service. DeepSeek documents account-level concurrency limits, but a limit on simultaneous requests is not the same as a monthly spending cap.
Protect the API key as part of cost control. The platform terms warn against exposing keys in browser or other client-side code because leaked credentials can lead to charges.
When comparing Flash with Pro, evaluate cost per successful task, not just the price per million tokens. A lower unit price does not automatically compensate for longer answers, more retries or additional review. Use the same test cases and quality requirements before choosing a model for a recurring workload.
Why another service may charge a different DeepSeek price
A third-party provider sets the prices for its own service. For example, OpenRouter’s DeepSeek listing presents provider-specific rates and offers. Those should not be substituted into a calculation for the first-party DeepSeek API.
Before comparing offers, check the exact model version, billing provider, input and output units, caching rules, discount conditions and any additional service fees. A shared “DeepSeek” label is not enough to establish that two offers are equivalent. A recurring charge from another app also does not, by itself, establish that DeepSeek sells the same subscription.
Self-hosting is a separate calculation. DeepSeek publishes model weights and local-inference documentation, but access to those files does not eliminate hardware, hosting, electricity or maintenance costs. Compare the complete operating cost with measured API spending—not “free weights” with a token price.