An LLM API pricing comparison is really a comparison of two numbers per model: what you pay to send text in, and what you pay to get text out. This page explains how that billing works, what a single question actually costs, and gives a dated LLM cost comparison for the five families on this site with the source for every figure.
Roughly how much does it cost a company when I ask an AI model one question?
You are billed by the token, not the question
A token is a chunk of text. Common short words are usually one token; longer or unusual words are split into several. For ordinary English, a useful rule of thumb is that 1,000 tokens is roughly 750 words. It is only a rule of thumb — punctuation, names and code all change the ratio.
Every request is billed twice over: input tokens for everything you send, and output tokens for everything the model writes. Output is almost always the more expensive of the two, often by a factor of three or four. That is why a model with a habit of writing 300 words when 60 would do is expensive in a way that is invisible in a price table.
What one question costs
Take a real example from this site. A short everyday question is a few dozen tokens in. The answers came back at very different lengths.
How long should I boil an egg for a soft yolk?
DeepSeek
3.6 s ✓Qwen
2.6 s ✓GLM
no answerMiniMax
2.3 s ✓GPT-OSS
2.6 s ✓At list prices of a dollar or two per million output tokens, a 200-word answer costs a fraction of a cent. That is why free tiers exist at all. It is also why they have daily limits: a fraction of a cent, multiplied by a lot of people asking a lot of questions, becomes a real number quickly.
An LLM API pricing comparison by family
These are list prices per million tokens from one large platform that hosts open models, read from its public price list on the date shown. Each family has several models at different sizes, so the figures are given as a band from the cheapest to the most expensive model of that family on that list. Prices change without notice; treat this table as a snapshot and follow the link for today’s numbers.
| Family | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| DeepSeek | $0.14 – $1.32source, checked 4 September 2026 | $0.28 – $3.96source, checked 4 September 2026 |
| Qwen | $0.15 – $2.00source, checked 4 September 2026 | $0.25 – $6.00source, checked 4 September 2026 |
| GLM | $0.15 – $1.40source, checked 4 September 2026 | $0.50 – $4.40source, checked 4 September 2026 |
| MiniMax | $0.30source, checked 4 September 2026 | $1.20source, checked 4 September 2026 |
| GPT-OSS | $0.15source, checked 4 September 2026 | $0.60source, checked 4 September 2026 |
Reading a price table without being misled
- Output dominates. Compare output prices first. A cheap input price on a verbose model is a false economy.
- Bands are wide. Within one family the largest model can cost ten times the smallest. “This family is cheap” is rarely a useful sentence.
- Cached input is cheaper. Several platforms bill repeated input at a large discount. If you send the same long instructions every time, that discount is the biggest lever you have.
- Batch is cheaper. Work that can wait is often billed at a lower rate.
- List price is not your price. Volume agreements, regions and dedicated capacity all change it.
- Free to you is not free. Any free tier is somebody paying these numbers.
A worked example
Suppose you send a 300-word question with a 200-word instruction, and get a 400-word answer. That is roughly 670 input tokens and 530 output tokens. At $0.50 per million input and $2.00 per million output, the request costs about $0.0014 — a seventh of a cent. A thousand such questions cost about $1.40.
The arithmetic is the useful part, not the numbers. Put your own volumes into it before believing anybody’s cost claim, including ours.
What the models say about their own cost
Roughly how much does it cost a company when I ask an AI model one question?
DeepSeek
6.6 s ✓It depends heavily on the model and the length of the question and answer, but a typical single interaction costs between $0.001 and $0.01 (a tenth of a cent to one cent).
Here’s the rough breakdown:
- Small/cheap models (e.g., a lightweight open-source model or a compact API like
ChatGPT-mini): ~$0.0005–$0.002 per question. - Mid-tier models (e.g., standard
ChatGPTorClaude Sonnet): ~$0.005–$0.02 per question, depending on length. - Top-tier/large models (e.g.,
ChatGPTorClaude Opuswith long outputs): ~$0.05–$0.20 per question.
What drives the cost:
- Input tokens (your question) – usually a few hundred tokens, costing fractions of a cent.
- Output tokens (the answer) – this is the bigger factor. A 500-word answer is ~700 tokens.
- Pricing per million tokens – e.g., $2–$15 per million input, $10–$60 per million output, depending on the model.
Real-world example:
- A 100-word question + a 300-word answer on ChatGPT: roughly $0.004 (0.4 cents).
- The same on ChatGPT (older, pricier): roughly $0.02 (2 cents).
Hidden costs: The company also pays for infrastructure, caching, and support, but those are usually amortized and small per query. For most companies, the marginal cost per question is well under a penny for common use cases.
Qwen
9.3 s ✓The cost varies significantly depending on whether the AI model is running on a massive consumer platform (like ChatGPT or Claude) or a specialized enterprise system, but for a single question, it is generally fractions of a cent.
Here’s a rough breakdown:
1. Consumer-Scale Models (e.g., ChatGPT, Gemini, Claude)
- Cost per query: $0.001 to $0.05 ($0.1¢ to 5¢).
- Why? These companies operate at massive scale. They use optimized hardware and efficient models. The cost includes electricity, hardware depreciation, and data center overhead.
- Note: Most free tiers are subsidized by paid subscriptions or enterprise contracts. If you’re using a free version, you aren’t directly paying, but the company still incurs this backend cost.
2. Enterprise/API Models (e.g., [provider name removed] API, Anthropic API)
If a company pays per token (word/character part), costs depend on:
- Model size: Smaller models (like ChatGPT-mini or Haiku) are much cheaper than large ones (like ChatGPT or Opus).
- Length of input/output: A short question with a short answer costs less than a long document analysis.
- Typical range: $0.0001 to $0.03 per simple Q&A interaction.
3. Specialized or On-Premise Models
- Cost per query: Highly variable, often higher initially due to lower utilization rates.
- Companies running their own servers pay fixed infrastructure costs regardless of usage, so the "cost per query" drops as more employees use it.
Key Factors That Influence Cost:
- Hardware Efficiency: Newer chips (like NVIDIA H100s) process queries
GLM
no answerNote that no family volunteers a current price with confidence. That is a good sign. Prices change more often than models are retrained, so a model quoting an exact figure would be quoting an old one.
Using this site costs you nothing; the daily limit exists because these numbers are real. Free AI chat without sign-up covers what free really means, the model policy explains how families are switched on and off, and the compare page lets you see how much text each family produces for the same question.
Frequently asked questions
What is a token, exactly?
A chunk of text produced by the model’s tokenizer. Roughly three quarters of a word in ordinary English, less for code and names. Models read, write and are billed in tokens rather than words.
Why is output more expensive than input?
Generating text requires a pass through the model for every token produced, while input can be processed in parallel. The price difference reflects the work.
Are these prices current?
They were read from the linked price list on the date shown in the table. Providers change prices without notice, so follow the link before relying on a figure.