DeepSeek API Pricing: Peak Rates, Off-Peak Savings, and Worked Cost Examples
By Nicho
DeepSeek API pricing depends on more than the model name. Cached input, generated output, and the time your requests run can change the bill substantially.
For DeepSeek Flash, CheapAIModel's displayed rates sit below DeepSeek's direct peak rates but above its direct off-peak rates. That distinction matters if you can schedule work. A blanket “30% cheaper than DeepSeek” claim misses the cheaper direct-access window.
Below is a dated price snapshot, followed by three calculated examples. These are budgeting scenarios, not measured application benchmarks or promises about how many successful tasks a token budget will buy.
Want to know how to pay? See DeepSeek API payment for PayPal, card top-ups, refunds and invoices.
Current DeepSeek Flash API prices
Prices checked on 26 October 2026. USD per one million tokens:
| Token category | DeepSeek direct: peak | DeepSeek direct: off-peak | CheapAIModel displayed rate |
|---|---|---|---|
| Input, cache miss | $0.30 | $0.15 | $0.209351 |
| Input, cache hit | $0.006 | $0.003 | $0.004187 |
| Output | $1.20 | $0.60 | $0.837404 |
Sources: DeepSeek Models & Pricing and CheapAIModel pricing.
DeepSeek's peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Other hours, weekends, and those holidays are off-peak. The official table also lists V4 Pro separately; do not apply Flash rates to Pro requests.
Before comparing providers, confirm the exact model and billing category. DeepSeek's direct API currently uses deepseek-flash, served by DeepSeek-V4.1-Flash-1017. CheapAIModel lists deepseek-v4-1-flash. Different route identifiers are not proof that every feature, model revision, or operating limit is identical. Direct API setup
How to calculate your DeepSeek API cost
Keep cached and uncached input separate:
Cost = (cache-miss input / 1,000,000 × miss rate) + (cache-hit input / 1,000,000 × hit rate) + (billable output / 1,000,000 × output rate).
Do not add cached input on top of total input. DeepSeek's Chat Completions usage schema defines prompt tokens as cache-hit plus cache-miss tokens. Its Responses usage schema also exposes reasoning tokens within the output breakdown. Use reported billable usage instead of counting only the final answer's visible words. Chat usage fields · Responses usage fields
The examples below include model-token charges only. They exclude separate tool charges, taxes where applicable, and external infrastructure. Actual usage must include retries and multi-step calls. A workload spanning both direct pricing windows needs separate subtotals rather than one assumed rate.
Example 1: A support workload
Assume 15 million uncached input tokens and 4 million output tokens:
- DeepSeek direct peak: 15 × $0.30 + 4 × $1.20 = $9.30.
- DeepSeek direct off-peak: $4.65.
- CheapAIModel: approximately $6.49.
These token counts do not guarantee a particular number of conversations. Longer history, retrieval context, and answer length change the cost per conversation. Use your application's usage records to estimate the number of chats this budget covers.
Example 2: Document summarization
Assume 200 million uncached input tokens and 20 million output tokens:
- DeepSeek direct peak: $84.00.
- DeepSeek direct off-peak: $42.00.
- CheapAIModel: approximately $58.62.
If the documents can wait, direct off-peak processing has the lowest token charge in this snapshot. If work must run immediately, compare the applicable current rates, operational requirements, and quality on your own documents.
Example 3: A cache-heavy coding workload
Assume 225 million input tokens, of which 180 million are cache hits and 45 million are misses, plus 22.5 million output tokens:
- DeepSeek direct peak: $41.58.
- DeepSeek direct off-peak: $20.79.
- CheapAIModel: approximately $29.02.
The 80% hit rate is an explicit scenario assumption, not an observed or guaranteed result. It also assumes the same token counts and cache classification across routes. Different caching behavior can change the comparison more than a small change in the advertised price.
All three examples are arithmetic using the displayed rates, rounded only at the final total. They are not latency, reliability, or task-success tests.
Is the DeepSeek API free?
The consumer chat app and the developer API are different products. DeepSeek introduced its app as free, while its API documentation describes usage charges deducted from an account balance. A free chat experience does not give your application a free API allowance. DeepSeek app announcement · API billing
Do not build a budget around a permanent free tier unless your account's current terms explicitly provide one. Check any promotional balance, expiry, and eligibility before treating it as available credit. Self-hosting is another option, but hardware, hosting, and operations still have costs.
CheapAIModel currently advertises prepaid billing, top-ups from $5, and no subscription. Check its current billing and refund terms before adding funds. A balance purchase is not the same as the token cost of a single request. Billing and prices
Three practical ways to control spending
1. Measure cache hits instead of assuming them
DeepSeek's cache matches repeated prompt prefixes on a best-effort basis. Keep stable instructions and reusable context at the beginning, then inspect prompt_cache_hit_tokens and prompt_cache_miss_tokens. Repeated-looking prompts do not guarantee a hit. Context caching guide
Track hit rates by provider and workload. A useful dashboard separates cached input, fresh input, and output, rather than showing only one impressive total-token number.
2. Tune reasoning for the task
The direct API enables thinking by default. Its OpenAI-format interface separates the thinking toggle from reasoning-effort control. Confirm the syntax for your chosen endpoint rather than copying settings from another provider. Thinking mode guide
Compare settings against a small set of tasks with explicit success criteria. A shorter answer is not a saving if it causes extra retries or fails the job. Use total cost per successful result as the decision metric.
3. Avoid sending unnecessary context
Inspect what the application sends on each turn: retrieved documents, tool definitions, previous messages, and repeated files. Remove irrelevant material carefully, then check that answer quality still meets your requirements. Set sensible output and application-level spending limits, but avoid cutting responses so aggressively that users must repeat the request.
Choose a route, then verify a small workload
For work that can run off-peak, start by comparing DeepSeek's direct off-peak price. For a multi-model application, evaluate whether CheapAIModel's shared access and billing fit your workflow, then compare total costs under the hours and cache behavior you actually expect.
Use a key issued by the provider you are calling, together with that provider's endpoint and model identifier. Do not mix an official DeepSeek key with a third-party endpoint.
Before moving production traffic, verify one representative request, the response format, supported features, usage fields, and the amount charged. This article does not claim that a production migration needs only three configuration changes.
Check current CheapAIModel prices before funding an account. For a broader model choice, see our LLM API pricing comparison, then validate your own workload before scaling it.