DEV Community

Cover image for How LLM Token Pricing Works: Input, Output, and Cache Explained
TokenPAPA
TokenPAPA

Posted on Originally published at doc.tokenpapa.ai

How LLM Token Pricing Works: Input, Output, and Cache Explained

#llm #pricing #api #tutorial

How LLM Token Pricing Works: Input, Output, and Cache Explained

Token pricing is the most misunderstood line on any LLM API bill. Developers often multiply the wrong number, forget that output is priced separately, or ignore caching entirely — and then wonder why their invoice is 5x the estimate.

This guide explains exactly how LLM token pricing works in 2026: what a token is, why input and output are billed differently, how context caching rewrites the math, and how to calculate a real cost. If you have been checking DeepSeek V4 API pricing per 1M tokens and getting confused by the fine print, this is the article for you.

What Is a Token, Anyway?

A token is the unit an LLM reads and writes. It is not a word and not a character — it is a chunk of text the model's tokenizer produces. Roughly:

  • English: 1 token ≈ 0.75 words (so 1M tokens ≈ 750,000 words)
  • Chinese: 1 token ≈ 1-2 characters
  • Code: 1 token ≈ 3-4 characters

Everything you send and everything the model returns is converted into tokens, and every token on both sides is billable. That is why token cost calculation starts with estimating token counts, not word counts — a 500-word English prompt is roughly 670 tokens, while a 500-character Chinese prompt can be 300-500 tokens depending on the tokenizer.

Input vs Output: Why Output Costs More

Every provider publishes two prices: input (prompt) and output (completion). Output is always more expensive — typically 3-10x — because generating tokens is far more compute-intensive than reading them.

The rule that saves developers the most money: always set max_tokens. One runaway reply can cost 20x a normal one, because every extra output token is billed at the highest rate on the bill.

The Price Table: Per 1M Tokens

To anchor your own DeepSeek V4 API pricing per 1M tokens and compare it with alternatives, here is the canonical table (input / output per 1M tokens):

Model Input /1M Output /1M Context Notes
Mimo V2.5 $0.08 $0.24 128K Cheapest absolute
DeepSeek V4 Flash $0.14 $0.42 128K Cost-effectiveness king
GPT-5.4 Mini $0.15 $0.60 128K OpenAI budget tier
Qwen 3.7 $0.20 $0.60 128K Coding + fallback
GPT-5.6 Luna $0.27 $2.70 1M Budget OpenAI tier
DeepSeek V4 Pro $0.28 $0.84 128K Best flagship value
GPT-5.6 Sol $13.50 $60.00 — Frontier flagship

The spread is the story: DeepSeek V4 Flash input is 96% cheaper than GPT-5.6 Sol ($0.14 vs $13.50). Same OpenAI-compatible format, one-line model= change, wildly different bill.

Context Caching: The 90% Discount

Most real applications re-send the same content on every request: a system prompt, product documentation, a long conversation history. Without caching, you pay full input price for those tokens every single time.

DeepSeek's automatic context caching changes that. When a prefix of your prompt repeats, it is served from cache at roughly 90% off the input rate — with no code changes and no opt-in. Long-context workloads that look expensive on paper often cost a fraction of the estimate once cache hits kick in.

A Real Calculation: What One Request Costs

Let's do a concrete token cost calculation for a typical request: 1K input + 0.5K output.

DeepSeek V4 Flash ($0.14 / $0.42):

  • Input: 1,000 / 1,000,000 × $0.14 = $0.00014
  • Output: 500 / 1,000,000 × $0.42 = $0.00021
  • Total: ~$0.00035 per request

GPT-5.6 Sol ($13.50 / $60.00):

  • Input: 1,000 / 1,000,000 × $13.50 = $0.01350
  • Output: 500 / 1,000,000 × $60.00 = $0.03000
  • Total: ~$0.04350 per request — 124x more

Scale that to a production workload of 100K requests/month: DeepSeek V4 Flash lands around $52/month, while GPT-5.6 Sol runs about $4,200/month. The $1 free credit at TokenPAPA covers roughly 2,800 requests on V4 Flash — enough to prototype a real product before you spend a cent.

FAQ

Q: Why does output cost more than input tokens?
A: Generating tokens is far more compute-heavy than reading them, so output is billed at 3-10x input — $0.14 vs $0.42 on DeepSeek V4 Flash, $13.50 vs $60 on GPT-5.6 Sol. Cap max_tokens to control it.

Q: How much does context caching save?
A: DeepSeek's automatic context caching cuts repeat input costs by about 90%. Re-sent system prompts, docs, and history get cached automatically — no code changes needed.

Q: How much does one API request actually cost?
A: A typical 1K-input + 0.5K-output request costs about $0.00035 on DeepSeek V4 Flash. The $1 free credit covers roughly 2,800 such requests.

Q: How does DeepSeek V4 API pricing per 1M tokens compare with other models?
A: DeepSeek V4 Flash ($0.14/$0.42) is 96% cheaper on input than GPT-5.6 Sol ($13.50/$60); Mimo V2.5 ($0.08/$0.24) is the absolute cheapest. All use the same OpenAI-compatible format.

Get Started

  1. Sign up at tokenpapa.ai — get $1 free credit
  2. Create your API key — OpenAI-compatible, no Chinese phone number needed
  3. Watch your bill shrink — one key for 30+ models, switch with a one-line model= change 
from openai import OpenAI

client = OpenAI(base_url="https://tokenpapa.ai/v1", api_key="your-key")

resp = client.chat.completions.create(
    model="deepseek-v4-flash",          # or "mimo-v2.5", "gpt-5.6-luna"
    messages=[
        {"role": "system", "content": "You are a concise assistant."},
        {"role": "user", "content": "Explain token pricing in one sentence."},
    ],
    max_tokens=100,                     # cap output to control cost
)
print(resp.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

Originally published at https://doc.tokenpapa.ai/en/docs/blog/llm-token-pricing-explained.

Top comments (0)