DeepSeek API Pricing Explained: Cost per Million Tokens Across 6 Models

DeepSeek^Ad's API bills by token, with three pricing tiers: input, cached input, and output. Price gaps between models can exceed 30x—pick the wrong model and costs spiral; pick the right one and chat workloads are nearly free. Based on a live pricing catalog from October 2026, this article covers the rates for all six DeepSeek models, the actual cost of two typical requests, and how caching further reduces your bill.

How DeepSeek API billing works

DeepSeek's billing is straightforward: the number of tokens consumed per request multiplied by the model's rate, priced per million tokens. No seat fees, no subscription plans, no minimum spend.

Of the three tiers, input and output are the baseline, while cached input is the key to savings. When a request's prefix (such as a long system prompt or a document) has already been processed by the API, the response marks those tokens as cache hits, billed at a cached rate far below the input price. For V4 Pro 0423, for example, input is $0.792 per million tokens while cached input is just $0.0338—a gap of more than 20x.

This also means that for applications repeatedly carrying the same context (such as customer service bots or document Q&A), the cache hit rate directly determines actual cost.

Tokens and prompt caching: A token is the smallest unit of text a model processes—roughly an English word or a fragment of a Chinese character. The API bills input tokens and output tokens separately. Prompt caching is a server-side optimization: when multiple requests share the same prefix (system prompts, documents, tool definitions, etc.), the server caches the already-computed KV state for that portion, so subsequent hits don't need recomputation and are billed at a lower rate. Caching typically matches prefixes exactly—once the prefix changes (even by inserting a sentence at the beginning), the rest can no longer hit the cache. So putting stable content first in the prompt and variable content last is a common way to improve hit rates.

Full price table for six models

The prices below come from the live catalog and are shown alongside DeepSeek's official list prices (USD per million tokens):

Model	Input	Cached	Output	Official Input	Official Cached	Official Output
DeepSeek V4 Pro 0813	$0.75	$0.0625	$1.188	$1.32	$0.044	$3.96
DeepSeek V4.1 Flash	$0.06	$0.0063	$0.18	$0.30	$0.006	$1.20
DeepSeek V4 Flash 0423	$0.0279	$0.0056	$0.0558	$0.30	$0.006	$1.20
DeepSeek V4 Flash 0731	$0.084	$0.0033	$0.09	$0.14	$0.0028	$0.28
DeepSeek V4 Flash Vision Exp	$0.132	$0.0042	$0.396	$0.30	$0.006	$1.20
DeepSeek V4 Pro 0423	$0.792	$0.0338	$1.92	$1.32	$0.044	$3.96

The lowest price in the table is DeepSeek V4 Flash 0423: $0.0279 input, $0.0558 output. At that price, processing a 50,000-token long document with 2,000 output tokens costs under 2 cents.

Interestingly, there are pricing differences within the Flash family. V4 Flash 0731's official list price is clearly lower than the 0423 version ($0.14 vs. $0.30 input), yet this channel's quote is actually higher ($0.084 vs. $0.0279). This shows that discount levels vary across versions—you can't simply assume "newer is cheaper."

"Official list price" vs. "channel quote": The two sets of numbers in the table represent two purchasing paths. Official list prices are the rates published on DeepSeek's website, typically at peak or standard tiers; channel quotes come from third-party API aggregation/relay providers, which obtain lower unit prices through bulk purchasing, credit packages, or promotions and resell at a discount. Channel prices may fluctuate with inventory and campaigns, and may offer bigger discounts on some versions while being more expensive on others (as with Flash 0731 above)—so you can't assume channel prices are always below official ones. When using a channel, also watch for stability, rate-limiting policies, and data compliance terms.

Actual cost of two typical requests

Consider two scenarios: a chat request consuming 2,000 input tokens and 500 output tokens; and a long-document request consuming 50,000 input tokens and 2,000 output tokens (both without cache hits).

Model	Chat/Official	Chat/This Channel	Long Doc/Official	Long Doc/This Channel
DeepSeek V4 Pro 0423	$0.0046	$0.0025	$0.0739	$0.0434
DeepSeek V4.1 Flash	$0.0012	$0.0002	$0.0174	$0.0034
DeepSeek V4 Flash 0423	$0.0012	$0.0001	$0.0174	$0.0015

The difference is obvious. For the same long-document request, V4 Pro 0423 at official list price costs $0.0739, while V4 Flash 0423 through this channel costs just $0.0015—nearly a 50x difference. Even within this channel, Pro and Flash differ by about 29x.

As for specific discounts: V4 Pro 0423 is 40% lower on input and 51% lower on output; V4.1 Flash is 80% lower on input and 85% lower on output; V4 Flash 0423 is 90% lower on input and 95% lower on output. The Flash family's discounts are notably larger than Pro's, which particularly benefits lightweight, high-volume tasks.

How cached input saves money

The caching rule is clear: when a request repeats a prefix the API has already seen, those tokens are billed at the cached rate. If a model has no separate cached rate, it falls back to the input rate.

This mechanism is most valuable in three scenarios. First, conversational apps with a fixed system prompt—each request carries the same instructions, so the cache hit rate can be high. Second, long-document analysis—when the same document is queried repeatedly, its tokens only need to be paid for at the cached rate. Third, agent-type applications—tool definitions and context templates reappear on every call.

Take V4 Flash 0423: the cached price of $0.0056 versus the input price of $0.0279 means that, once cached, that portion of cost drops to about one-fifth. In long-context scenarios, this saving is often more significant than model choice itself.

How to call it

All models share the same calling method—just point the SDK's base URL to the provider's address and use the corresponding API key:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.uplinkswitch.com/v1",
    api_key="sk-uplink-live-...",
)

completion = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Summarize this in one line."}],
)
print(completion.choices[0].message.content)

Because the interface is compatible with the OpenAI SDK, existing code generally only needs its base URL and model ID swapped, keeping migration costs low.

OpenAI-compatible interfaces and base_url: The OpenAI SDK lets you point requests to any server implementing the same HTTP protocol via the base_url parameter—such servers must provide endpoints like /v1/chat/completions and accept the same request body structure (model, messages, temperature, etc.). Since DeepSeek and most aggregation providers are compatible with this protocol, migration typically requires changing only three things—base_url, api_key, and model—while message construction, streaming parsing, function calling, and other logic in your business code can be reused as-is. Note that providers may differ slightly in model ID naming, supported parameter subsets, and error codes, so run a smoke test after switching.

Model selection advice and FAQs

Flash or Pro? The original advice is practical: use Flash for chat, summaries, and agent subtasks—cheap enough that you don't need to count pennies; use Pro for code and reasoning tasks, where Flash tends to make mistakes. The safest approach is to run the same prompt on both models and compare output quality before deciding.

Is there a subscription plan? No. Top-ups start at $10, and each top-up locks the price for 30 days. That means prices roll with your top-up cycle, so for long-term use you should watch for rate changes at renewal.

How do I estimate my own costs? The most direct way is to export your actual usage and recalculate using the rates above. For applications with existing bills, this is far more reliable than estimating by feel.

One final reminder: the official list prices in the table are taken from DeepSeek's pricing page at peak rates, while channel quotes come from the live catalog. Both will change, so for budgeting, rely on the price at the time of actual calls rather than treating any single point-in-time number as a long-term commitment.