DeepSeek API pricing 2026 is based on token consumption rather than fixed monthly subscription tiers. The final cost depends on the selected model, the number of uncached input tokens, any input tokens served from context cache, generated output tokens and whether the request is processed during peak or off-peak hours.

The current API lineup includes DeepSeek V4 Flash and DeepSeek V4 Pro. Flash is the lower-cost option for routine and high-volume workloads, while Pro costs more and is intended for complex reasoning, coding and agent-based tasks.

DeepSeek API Pricing at a Glance

The following rates are listed in US dollars per one million tokens.

Model Billing category Off-peak rate Peak rate
DeepSeek V4 Flash Uncached input $0.22 $0.44
DeepSeek V4 Flash Cached input $0.007 $0.014
DeepSeek V4 Flash Output $0.66 $1.32
DeepSeek V4 Pro Uncached input $0.66 $1.32
DeepSeek V4 Pro Cached input $0.022 $0.044
DeepSeek V4 Pro Output $1.98 $3.96

DeepSeek’s official Models and Pricing page should be checked before making a purchasing or deployment decision because API rates, model aliases and billing rules can change.

How DeepSeek API Billing Works

DeepSeek measures usage in tokens. A token is a small unit of text processed or generated by the model. It may represent a whole word, part of a word, punctuation or another piece of encoded content.

Every request can include three separately priced token categories:

  • Uncached input tokens
  • Cached input tokens
  • Output tokens

The API does not charge one flat price for the entire request. Each category is multiplied by its applicable rate and then added to the total.

The basic calculation is:

Request cost = uncached input cost + cached input cost + output cost

More precisely:

Cost = (uncached input tokens ÷ 1,000,000 × uncached rate) + (cached input tokens ÷ 1,000,000 × cached rate) + (output tokens ÷ 1,000,000 × output rate)

The selected peak or off-peak rate must be used consistently for the request.

Input Tokens

Input tokens include the content sent to the model, such as:

  • System instructions
  • User prompts
  • Conversation history
  • Source documents
  • Code
  • Tool definitions
  • Tool results returned to the model

A request containing a 10,000-token document and a 500-token instruction has approximately 10,500 input tokens before any additional message formatting or conversation history is considered.

Long conversations can become more expensive because the DeepSeek Chat Completions API is stateless. An application normally sends the relevant conversation history again with each new request. Repeated portions may qualify for context-cache pricing, but they still count toward context usage.

The DeepSeek API guide explains how requests, model names, authentication and multi-turn input work together.

Cached Input Tokens

DeepSeek automatically uses context caching for repeated input prefixes. When part of a new request fully matches a prefix previously stored in the cache, those matched tokens can be billed at the much lower cached-input rate.

For example, an application may repeatedly send:

  • The same system instructions
  • A large reference document
  • A stable product catalog
  • An unchanged codebase summary
  • A standard set of tool definitions
  • Previous messages from the same conversation

If the matching prefix is available in cache, DeepSeek reports the reused portion as cached input rather than ordinary uncached input.

Cached input is significantly less expensive. During off-peak hours, one million cached input tokens cost $0.007 with V4 Flash, compared with $0.22 for one million uncached input tokens.

Caching does not make the whole request free. New instructions, changed prefixes and generated output are still billed at their respective rates.

Output Tokens

Output tokens are generated by the model. They include the final response and can also reflect additional model output associated with reasoning-intensive requests.

Short classifications or extracted fields may use only a small number of output tokens. Detailed reports, code generation, long explanations and agent workflows can produce substantially more.

Output tokens cost more than input tokens for both V4 models. Controlling unnecessary output is therefore an important part of managing API spending.

Applications can reduce avoidable output by specifying:

  • The required response length
  • An exact output schema
  • The number of requested alternatives
  • Whether explanations are necessary
  • Whether thinking mode is appropriate
  • A sensible output-token limit

Reasoning effort should match the task. The DeepSeek thinking mode guide explains how low, high and max effort affect model behavior.

Peak and Off-Peak Pricing

DeepSeek uses time-based pricing during the working week.

Peak hours are:

  • 01:00–04:00 UTC
  • 06:00–10:00 UTC
  • Monday through Friday

All remaining periods use off-peak pricing. This includes:

  • Weekday hours outside the two peak windows
  • The gap between the two weekday peak windows
  • Saturdays
  • Sundays

Off-peak rates are half of the corresponding peak rates.

A request made at 05:00 UTC on a weekday falls within off-peak time. A request made at 07:00 UTC on the same day uses peak pricing. The relevant time is the API billing schedule in UTC, not the developer’s local clock.

Teams scheduling batch processing should convert these UTC windows carefully, especially in regions affected by daylight-saving time.

DeepSeek V4 Flash Pricing

DeepSeek V4 Flash is the economical model in the V4 family.

Its off-peak rates per one million tokens are:

  • $0.22 for uncached input
  • $0.007 for cached input
  • $0.66 for output

Its peak rates are:

  • $0.44 for uncached input
  • $0.014 for cached input
  • $1.32 for output

Flash is a practical starting point for:

  • Classification
  • Data extraction
  • Summarization
  • Routine writing
  • Support-response drafts
  • Code explanation
  • Predictable transformations
  • High-volume tool calls
  • Interactive applications where latency matters

It still supports thinking mode, coding and agent workflows. Its lower price does not mean it is restricted to basic chat.

DeepSeek V4 Pro Pricing

DeepSeek V4 Pro is the higher-capability and higher-cost model.

Its off-peak rates per one million tokens are:

  • $0.66 for uncached input
  • $0.022 for cached input
  • $1.98 for output

Its peak rates are:

  • $1.32 for uncached input
  • $0.044 for cached input
  • $3.96 for output

Pro is designed for work such as:

  • Complex software engineering
  • Repository-level coding
  • Difficult debugging
  • Multi-step technical analysis
  • Advanced mathematical reasoning
  • Agent workflows involving several tools
  • Tasks where a weak response would create significant human rework

The DeepSeek V4 comparison provides more detail about the capabilities, context window and intended uses of both models.

V4 Flash vs V4 Pro Cost Comparison

For uncached input and output, V4 Pro costs three times as much per token as V4 Flash under both pricing periods.

Cost factor V4 Flash V4 Pro
Uncached input Baseline 3× Flash
Cached input Baseline Slightly more than 3× Flash
Output Baseline 3× Flash
Off-peak discount 50% below peak 50% below peak

The price difference does not establish which model is more economical for every task.

A Pro response that solves a difficult problem correctly in one attempt may cost less overall than several failed Flash attempts followed by extensive human correction. Conversely, using Pro for routine extraction or rewriting can increase spending without producing meaningful additional value.

A common routing strategy is to send routine requests to Flash and escalate only difficult cases to Pro.

Cost Example 1: A Short V4 Flash Request

Suppose one V4 Flash request uses:

  • 100,000 uncached input tokens
  • No cached input
  • 20,000 output tokens

During off-peak time:

  • Input: 0.1 × $0.22 = $0.022
  • Output: 0.02 × $0.66 = $0.0132
  • Total: $0.0352

The same request during peak time costs:

  • Input: 0.1 × $0.44 = $0.044
  • Output: 0.02 × $1.32 = $0.0264
  • Total: $0.0704

The peak total is exactly twice the off-peak total.

Cost Example 2: V4 Pro With Cached Input

Consider a V4 Pro request containing:

  • 400,000 cached input tokens
  • 100,000 uncached input tokens
  • 50,000 output tokens

During off-peak time:

  • Cached input: 0.4 × $0.022 = $0.0088
  • Uncached input: 0.1 × $0.66 = $0.066
  • Output: 0.05 × $1.98 = $0.099
  • Total: $0.1738

During peak time:

  • Cached input: 0.4 × $0.044 = $0.0176
  • Uncached input: 0.1 × $1.32 = $0.132
  • Output: 0.05 × $3.96 = $0.198
  • Total: $0.3476

This example shows why cached and uncached input must be calculated separately.

Cost Example 3: Comparing Flash and Pro

Assume a workload uses:

  • One million uncached input tokens
  • 250,000 output tokens
  • Off-peak processing

With V4 Flash:

  • Input: 1 × $0.22 = $0.22
  • Output: 0.25 × $0.66 = $0.165
  • Total: $0.385

With V4 Pro:

  • Input: 1 × $0.66 = $0.66
  • Output: 0.25 × $1.98 = $0.495
  • Total: $1.155

For this identical token volume, Pro costs three times as much as Flash.

This is a mathematical illustration, not evidence that both models will produce equal-quality results. Actual model choice should account for accuracy, completion rate and human-review time.

Cost Example 4: A Monthly Application Workload

Suppose an application processes 10,000 requests. Each request contains:

  • 2,000 input tokens
  • 500 output tokens

Monthly usage is therefore:

  • 20 million input tokens
  • 5 million output tokens

If all input is uncached and all processing uses V4 Flash off-peak:

  • Input: 20 × $0.22 = $4.40
  • Output: 5 × $0.66 = $3.30
  • Total: $7.70

If 15 million input tokens qualify for cache hits and only five million remain uncached:

  • Cached input: 15 × $0.007 = $0.105
  • Uncached input: 5 × $0.22 = $1.10
  • Output: 5 × $0.66 = $3.30
  • Total: $4.505

The lower total is a calculated scenario based on the stated token volumes. It is not a guaranteed saving because DeepSeek describes context caching as a best-effort system.

Running the original fully uncached workload on V4 Pro off-peak would cost:

  • Input: 20 × $0.66 = $13.20
  • Output: 5 × $1.98 = $9.90
  • Total: $23.10

How DeepSeek Context Caching Works

DeepSeek context caching is enabled automatically. Developers do not need to purchase a separate cache plan or activate a special subscription.

When requests share an identical prefix, DeepSeek may store and reuse that prefix. A later request can receive a cache hit when it fully matches a persisted prefix unit.

A simplified sequence could look like this:

  1. An application sends a system prompt and a large document.
  2. DeepSeek processes the request and constructs cache data.
  3. A later request starts with the same system prompt and document.
  4. The matching portion may be reported as cached input.
  5. Only the new or unmatched portion receives the uncached-input rate.
  6. The model still generates a new output, billed at the output rate.

Caching affects input processing only. It does not reuse a previous answer or guarantee identical output.

The API response includes two useful fields:

  • prompt_cache_hit_tokens
  • prompt_cache_miss_tokens

These fields allow developers to measure how much input received each rate. The dedicated DeepSeek context-caching guide explains prefix persistence and cache-hit behavior.

Important Context-Cache Limitations

Context caching can reduce costs, but several limitations matter:

  • The first request cannot rely on an existing cache unless the prefix was previously persisted.
  • Similar wording is not necessarily enough; the reusable prefix must match.
  • Changing early instructions can prevent later content from matching.
  • Cache construction takes time.
  • A cache hit is not guaranteed.
  • Unused cache entries may be cleared after hours or days.
  • New input remains subject to the uncached rate.
  • Output is always generated again and billed normally.

Applications should use actual API usage data rather than assuming a fixed cache-hit percentage.

Improving Cache-Hit Potential

A stable prompt structure can make repeated prefixes easier to reuse.

Place relatively fixed content first:

  1. Stable system instructions
  2. Tool definitions
  3. Shared reference material
  4. Existing conversation history
  5. New task-specific instructions

Avoid changing punctuation, whitespace or wording in the stable prefix without a reason. Dynamic information placed near the beginning of every prompt may reduce the amount of reusable content.

Prompt design should still prioritize accuracy. Moving essential instructions merely to chase cache savings can create weaker results and higher human-review costs.

How Thinking Mode Affects Cost

Thinking mode can use additional tokens while the model works through a problem. Higher reasoning effort may therefore increase latency and output-side usage.

DeepSeek provides low, high and max reasoning-effort levels. Max effort should not be enabled automatically for every request.

A cost-aware routing policy might use:

  • Non-thinking mode for formatting and simple extraction
  • Low effort for straightforward analysis
  • High effort for ordinary coding and agent tasks
  • Max effort for genuinely difficult problems

The correct comparison is not only the price of one request. Teams should measure how often each configuration produces an acceptable result without retries or extensive correction.

Additional Costs Outside the Token Rate

The DeepSeek token table covers model inference. A complete application may also incur costs for:

  • Cloud hosting
  • Databases
  • Vector storage
  • Search services
  • Monitoring and logging
  • Network traffic
  • File processing
  • Developer time
  • Human review
  • Third-party agent platforms

Using DeepSeek through another API provider or marketplace may result in different prices, fees, model revisions or caching rules. The rates in this article apply to the official DeepSeek API rather than every third-party service offering a DeepSeek model.

How to Estimate a Real Budget

A useful API budget should be based on measured usage.

Collect the following information:

  • Requests per day
  • Average uncached input tokens
  • Average cached input tokens
  • Average output tokens
  • Percentage processed during peak hours
  • Model used for each task
  • Retry rate
  • Tool-call rounds per task
  • Reasoning-effort level
  • Expected growth in request volume

Calculate Flash and Pro separately. If an application routes only difficult cases to Pro, estimate the percentage of requests likely to be escalated.

Adding a reasonable operational buffer can account for unusually long conversations, retries, traffic spikes and lower-than-expected cache performance.

Ways to Control DeepSeek API Costs

Use Flash for Suitable Tasks

Begin with V4 Flash for routine work. Move a task to Pro only when testing shows a meaningful quality advantage.

Schedule Flexible Work Off-Peak

Run document processing, batch summarization, indexing or evaluation jobs outside the weekday peak windows when business requirements allow.

Keep Repeated Prefixes Stable

Stable instructions and shared documents can improve the opportunity for cache hits.

Limit Unnecessary Conversation History

Send the context required for the current task rather than carrying unrelated earlier messages into every request.

Control Output Length

Request concise formats where a long explanation is not needed. Output tokens are more expensive than uncached input tokens.

Match Reasoning Effort to Complexity

Maximum reasoning can be valuable for difficult work but wasteful for predictable transformations.

Monitor Usage Fields

Record total input, cached input, uncached input and output tokens for each task type. Estimates based only on character counts can miss important usage patterns.

Test Total Cost per Successful Task

A cheaper request is not economical if it frequently fails. Measure cost together with accuracy, retries and human correction time.

How to Verify the Official Price

API pricing can change, so published articles should not be the final source for a live budget.

Use this verification process:

  1. Open the official DeepSeek Models and Pricing documentation.
  2. Confirm that the model names are deepseek-v4-flash and deepseek-v4-pro.
  3. Check the uncached input, cached input and output rates separately.
  4. Verify the peak windows in UTC and confirm whether they apply only on weekdays.
  5. Review the official DeepSeek API changelog for pricing or model updates.
  6. Inspect prompt_cache_hit_tokens and prompt_cache_miss_tokens in real API responses.
  7. Compare calculated usage with the balance or billing records in the official developer platform.
  8. Recheck third-party pricing separately if DeepSeek is accessed through another provider.

This method avoids relying on an outdated screenshot, search snippet or unofficial price comparison.

Frequently Asked Questions

How is the DeepSeek API priced?

The official API is billed according to token usage. Uncached input, cached input and generated output have separate rates.

Does DeepSeek offer fixed API subscription plans?

The official V4 API pricing described here is token-based. Fixed monthly plans offered by a separate interface, application or third-party provider should not be presented as official DeepSeek API plans.

What is the cheapest current DeepSeek V4 model?

V4 Flash has the lower per-token rates. V4 Pro costs roughly three times as much for ordinary uncached input and output.

What is the difference between cached and uncached input?

Uncached input must be processed normally. Cached input matches a previously persisted request prefix and receives a lower input rate.

Is context caching automatic?

Yes. DeepSeek states that context caching is enabled by default. Cache hits are best-effort and are not guaranteed.

When does peak pricing apply?

Peak pricing applies Monday through Friday from 01:00–04:00 UTC and 06:00–10:00 UTC.

Are weekends off-peak?

Yes. Under the published schedule, Saturdays and Sundays fall outside the weekday peak periods.

How much cheaper is off-peak pricing?

Each listed off-peak rate is 50% of the corresponding peak rate.

Does a one-million-token context request cost the same every time?

No. The cost depends on how many tokens are actually used, how much input receives a cache hit, which model generates the response, how many output tokens are produced and when the request is processed.

Does thinking mode cost extra?

There is no separate thinking-mode subscription fee. However, reasoning can increase token usage, which may increase the total request cost.

Where should current DeepSeek pricing be checked?

Use the official DeepSeek Models and Pricing documentation and confirm the usage shown in the developer platform. Third-party articles should be treated as explanatory resources rather than live billing authorities.

Final Assessment

DeepSeek API pricing 2026 is a token-based system with separate rates for uncached input, cached input and output. V4 Flash provides the lower-cost route for routine or high-volume workloads, while V4 Pro charges more for stronger performance on complex reasoning, coding and agent tasks.

Peak requests cost twice as much as off-peak requests. Automatic context caching can substantially reduce the input portion of repeated workloads, but cache hits are not guaranteed and generated output is still billed normally.

Reliable cost planning requires actual token measurements, correct UTC scheduling, model routing and regular verification against DeepSeek’s official pricing page. These factors provide a more dependable budget than invented subscription tiers, anonymous competitor comparisons or unsupported saving claims.