# AI Model Pricing Per Million Tokens, Explained

AI model prices are quoted per million tokens, with one rate for input and a higher one for output. To estimate what a message costs, multiply the tokens you send by the input rate, multiply the tokens you get back by the output rate, and add the two. This guide shows the arithmetic with real list prices from Quad Chat's model catalog, and explains how that differs from a plan that bills shared credits.

> **Direct answer:** A typical chat message of 1,500 input tokens and 700 output tokens costs about $0.0005 on a lightweight model such as GPT-6 Luna and about $0.02 on Claude Opus 5.5, a 40 times gap at list prices. Per-token billing exposes that gap on every request. A flat plan with shared credits, like Quad's, hides the arithmetic but keeps the same logic: cheaper models use less of your allowance.

## What a token is

A token is the unit a model reads and writes. It is a chunk of text, often a short word or part of a longer one. As a rough rule of thumb, a token is about three-quarters of an English word, but the ratio varies by model, by language and by content such as code. Providers count tokens, not words or messages, which is why prices are quoted that way.

"Per million tokens" is simply the unit that keeps the numbers readable. A price of $2 per million input tokens means 1,000 input tokens cost $0.002.

## Input, output and why they are priced differently

Every request has two sides:

- **Input tokens** are everything you send: your message, the conversation so far, attached files and any system instructions.
- **Output tokens** are what the model writes back. For reasoning models this can include thinking tokens that you do not always see.

Output is priced higher because generating tokens takes more compute than reading them. In Quad's catalog, the output rate is five times the input rate for GPT-6.1 Sol, Claude Sonnet 5.5 and Claude Opus 5.5, six times for Gemini 3.1 Pro, and about two times for DeepSeek V4 Flash.

## The formula

Cost per request = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price).

For Claude Opus 5.5 at $4 input and $20 output per million tokens, a message with 1,500 input and 700 output tokens costs:

- Input: 1,500 ÷ 1,000,000 × $4 = $0.006
- Output: 700 ÷ 1,000,000 × $20 = $0.014
- Total: $0.02

## What a typical message costs on ten models

The table uses the same assumption as Quad's pricing page, about 1,500 input tokens and 700 output tokens, and the API list prices recorded in the model catalog. Each model has its own page in the [model directory](/models) with its context window, plan access and list price.

| Model | Input per 1M | Output per 1M | Cost per message | Messages per $1 | Lowest Quad plan |
| --- | ---: | ---: | ---: | ---: | --- |
| DeepSeek V4 Flash | $0.076 | $0.153 | $0.00022 | about 4,500 | Free |
| GPT-6 Luna | $0.10 | $0.50 | $0.0005 | 2,000 | Free |
| Grok 4.1 Fast | $0.20 | $0.50 | $0.00065 | about 1,540 | Free |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.01 | 100 | Pro |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.01 | 100 | Pro |
| Gemini 3.1 Pro | $2.00 | $12.00 | $0.0114 | about 88 | Pro |
| Kimi K3 | $3.00 | $15.00 | $0.015 | about 67 | Go |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.02 | 50 | Ultra |
| GPT-6 Astra | $10.00 | $50.00 | $0.05 | 20 | Pro |
| Claude Fable 5 | $10.00 | $50.00 | $0.05 | 20 | Ultra |

Read it as scale, not as a ranking of quality. A message on GPT-6 Astra costs about a hundred times a message on GPT-6 Luna, yet for a quick rewrite the cheaper model is often enough. Choosing the right tier for the task is the biggest cost lever you have. See [GPT-6](/models/gpt-6) and [Claude Opus 5.5](/models/claude-opus-5-5) for the full variant lists.

## Context length is where bills jump

The 1,500-token message above is small. The conversation history and attachments are also input, and they are sent again with each turn. Attach a long document and the input side dominates.

A 100,000-token document and a 2,000-token answer costs:

- Claude Sonnet 5.5: 100,000 × $2 ÷ 1M = $0.20 input, plus 2,000 × $10 ÷ 1M = $0.02 output, so about $0.22.
- GPT-6 Luna: 100,000 × $0.10 ÷ 1M = $0.01 input, plus 2,000 × $0.50 ÷ 1M = $0.001 output, so about $0.011.

Many providers offer discounts for cached input or batch jobs. The table ignores them, so treat it as a plain list-price comparison.

## How per-token billing compares with a flat plan

There are three common ways to pay for model access:

| Approach | You pay | Cost behavior | Fits |
| --- | --- | --- | --- |
| Direct API or a router | Per token, usually prepaid | Exact, but unbounded and variable | Developers shipping software |
| Bring your own keys | Per token to each provider, plus the app | Exact, spread across several bills | Technical users who want control |
| Flat plan with shared credits | A fixed monthly price | Predictable, capped by the plan | People who use several models for daily work |

Quad uses the third approach. Every plan includes a monthly allowance of credits that all models share. Cheaper models use fewer credits, and the model picker shows an estimate before you send. The list prices above are what providers charge; Quad's credits are its own unit, so a credit is not a fixed number of tokens. [How AI credits work](/blog/how-ai-credits-work-multi-model-app) explains the plan numbers: for example, Pro includes 1,200 credits and 1,000 web searches a month, and the pricing page estimates about 800 Claude Sonnet 5.5 messages on Pro using the message size above.

For a closer look at the other two approaches, see [Quad vs OpenRouter](/blog/quad-chat-vs-openrouter) and [Quad vs TypingMind](/blog/quad-chat-vs-typingmind).

## Practical ways to spend less

1. **Default to a lighter model** and escalate only when the answer falls short.
2. **Keep context tight.** Start a fresh conversation when the old one no longer matters, because history is re-sent as input.
3. **Ask for the length you need.** Output tokens cost more than input tokens.
4. **Reserve flagship models** for work where a better answer is worth the higher price.
5. **Compare before you commit.** When two models cost very different amounts, [comparing them side by side](/blog/how-to-compare-ai-models-side-by-side) on a real task shows whether the premium buys anything.

## Summary

Per-million-token pricing is two rates and a formula. The gap between the cheapest and most expensive models is large, tokens in long documents add up quickly, and the right choice depends on the task. A flat plan trades exact metering for a predictable bill and one pool of credits across every model.

**[Try Quad Chat free ->](/)**

List prices were read from Quad's model catalog on October 5, 2026. Providers change prices, so check the model's page or the provider before relying on a figure.

See also: [How AI credits work in a multi-model app](/blog/how-ai-credits-work-multi-model-app) | [Quad pricing](/pricing) | [One subscription for ChatGPT, Claude and Gemini](/blog/one-subscription-chatgpt-claude-gemini)