What Is 1 Million Tokens in AI Pricing?
One million tokens is approximately 750,000 words of English text, roughly 1,500 pages in a standard document, or about 3–4 typical novels. In code, it's around 200,000–300,000 lines depending on the language and verbosity.
LLM API providers price their services per million tokens because it creates human-scale numbers. Instead of saying "$0.0000025 per token," providers say "$2.50 per million input tokens"—much easier to reason about when you're processing thousands or millions of tokens daily.
Understanding what a million tokens represents helps you estimate costs, plan budgets, and optimize your AI applications. Whether you're processing customer support tickets, generating reports, or analyzing documents, knowing token volumes matters.
Visualizing 1 Million Tokens
Token counts vary by content type, language, and tokenizer. Here's what 1 million tokens looks like across different content:
| Content Type | Approximate Volume | Real-World Example |
|---|---|---|
| English Prose | ~750,000 words | 3–4 typical novels (250–300 pages each) |
| Document Pages | ~1,500 pages | A complete PhD dissertation or technical manual |
| Code (Python) | ~200,000–250,000 lines | A medium-sized application codebase |
| Code (JavaScript) | ~180,000–220,000 lines | React app + dependencies (minified) |
| JSON Data | ~300,000–400,000 lines | Large API response dataset |
| Chinese Text | ~500,000 characters | 2–3 typical novels (Chinese uses more tokens) |
| Conversation Logs | ~5,000–10,000 exchanges | A month of customer support chat history |
Token Density Varies
The same content produces different token counts across models:
- GPT-4o (o200k_base): Most efficient English tokenizer, baseline reference
- GPT-4 (cl100k_base): ~10–15% more tokens than GPT-4o
- Claude (Anthropic tokenizer): ~50% more tokens for English, ~70% more for code
- Gemini (SentencePiece): ~20–30% more tokens than GPT-4o
- Chinese-optimized (GLM, Qwen): More efficient for CJK languages, similar to GPT for English
This means 1 million tokens on GPT-4o might be only ~650,000 tokens on Claude for the same text. Always count per-provider.
How Much Does 1 Million Tokens Cost?
Costs vary dramatically by model and whether tokens are input (reading) or output (generating):
| Model | Input Cost | Output Cost | Combined (50/50) |
|---|---|---|---|
| Gemini 3.8 Flash | $0.075 | $0.30 | $0.19 |
| GPT-4o mini | $0.15 | $0.60 | $0.38 |
| DeepSeek V4.1 | $0.27 | $1.10 | $0.69 |
| Claude 5.5 Haiku | $0.80 | $4.00 | $2.40 |
| GPT-4o | $2.50 | $10.00 | $6.25 |
| Claude 5.5 Sonnet | $3.00 | $15.00 | $9.00 |
| Grok Beta | $5.00 | $15.00 | $10.00 |
| GPT-6 Sol | $12.00 | $60.00 | $36.00 |
The range is massive: from $0.19 (Gemini 3.8 Flash) to $36.00 (GPT-6 Sol) for a balanced workload. Choosing the right model for your use case can reduce costs by 100× or more.
Practical Examples: When You Hit 1 Million Tokens
Chatbot Processing 10,000 Conversations
A customer support chatbot handling 10,000 conversations per day, averaging 100 tokens per exchange (50 input + 50 output), processes 1 million tokens daily. At GPT-4o pricing ($2.50 input + $10.00 output), that's $6.25/day or ~$190/month.
Document Summarization Pipeline
Summarizing 100 research papers, each 10,000 tokens, into 200-token summaries: 1M input tokens + 20K output. Using Claude 5.5 Sonnet: (1M × $3.00) + (20K × $15.00/M) = $3.00 + $0.30 = $3.30 total.
Code Analysis Tool
Analyzing 5,000 lines of Python code (typically ~15K tokens) for 67 files totals ~1M tokens input. Using GPT-4o mini for analysis with minimal output: $0.15 input + ~$0.05 output = ~$0.20 total for the entire codebase analysis.
Data Extraction from Invoices
Processing 2,000 invoice images via GPT-4V (Vision), each image ~500 tokens: 1M tokens total. With structured JSON output (~50 tokens per invoice), cost is (1M × $2.50) + (100K × $10.00/M) = $2.50 + $1.00 = $3.50 for the batch.
Frequently Asked Questions
How long does it take to process 1 million tokens?
For input, most models process 1M tokens in under 10 seconds. For output generation at typical speeds (50–100 tokens/second), generating 1M tokens would take 3–6 hours of continuous generation. In practice, you'd parallelize across many requests.
Can I process 1 million tokens in a single request?
Most models have context windows under 200K tokens (GPT-4o: 128K, Claude 5.5: 200K). Gemini 4 Argon supports up to 1M tokens in a single request, but you'd still need to stay within that limit for input + output combined.
Is there a volume discount for millions of tokens?
Some providers offer enterprise pricing at scale. OpenAI's Batch API provides 50% discounts for async processing. DeepSeek offers off-peak pricing. Check with providers directly for custom pricing above certain thresholds.
How many tokens do I use per month in a typical app?
Highly variable. A small chatbot might use 10–50M tokens/month ($20–500). A large-scale content generation platform might use billions of tokens monthly ($10K–$100K+). Use our calculator to estimate your specific usage.
Are cached tokens counted toward my million?
Yes, but they're billed at reduced rates. Anthropic's cache reads cost ~10% of regular input pricing. So 1M cached tokens might cost $0.30 instead of $3.00, but they still count as 1M tokens in volume.