# What does AI really cost: subscriptions, API, tokens and credits?

Date: 2026-10-28 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-ai-really-costs-2026/

**Answer:** For one person chatting, AI costs a flat subscription, typically $20 a month in October 2026, with usage limits rather than a meter. For anything you build, you pay per token through an API, from about $0.10 to $10 per million input tokens depending on the model, and the same job can cost 100 times more on a flagship model than on a small one.

AI is sold in two very different ways, and most confusion about cost comes from mixing them up. A **subscription** gives a person access to an app such as ChatGPT or Claude for a fixed monthly price, with limits on how much you can use it. An **API** (application programming interface) lets software call a model directly, and you pay for exactly what you use, measured in tokens. On top of both sit **credits**: prepaid or included amounts of money that the API draws from.

The site's earlier guides explain per-token pricing and why prices keep falling. This update collects the numbers as they stand at the end of October 2026 and the newer details that change the real bill.

## Subscriptions: flat price, soft limits

For individuals, a typical entry plan costs about $20 a month. Anthropic's Claude Pro, for example, is $20 billed monthly, or $17 a month on an annual plan, and includes the Claude Code coding tool, which shares the same usage limits as chat. Heavier users pay more: Claude Max starts at $100 a month for five or twenty times Pro's usage. Free tiers exist, but with fewer models and tighter limits.

A newer twist is **included API credits**. Claude's Max plans now come with $100 or $200 a month to spend on the Claude API for your own projects. Read the fine print: unused credits vanish when each billing period ends, do not carry over, and do not raise your chat limits.

## API prices: tokens in, tokens out

A **token** is a piece of text, often a word or part of one. APIs charge separately for the tokens you send (input) and the tokens the model writes (output), and output usually costs five times more. Standard prices per million tokens as of October 2026:

| Model | Input | Output |
|---|---|---|
| Claude Haiku 5.5, GPT-6 Luna | $0.10 | $0.50 |
| GPT-6.1 Terra | $1 | $5 |
| Claude Sonnet 5.5, GPT-6.1 Sol | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Astra, Claude Fable 5.1 | $10 | $50 |

One hidden factor: the same text does not produce the same number of tokens on every model. Anthropic notes that its Claude 4.7 and later models use a newer tokenizer that turns the same text into roughly 30% more tokens, so a price per token is only comparable once you count tokens on the model you will use.

## A worked example: 1,000 emails

Say you want short summaries of 1,000 customer emails. Each email plus instructions is about 1,500 tokens, and each summary about 200. That is 1.5 million input tokens and 0.2 million output tokens.

| Model | Input cost | Output cost | Total |
|---|---|---|---|
| Haiku 5.5 or Luna | $0.15 | $0.10 | $0.25 |
| Sonnet 5.5 or GPT-6.1 Sol | $3.00 | $2.00 | $5.00 |
| Astra or Fable 5.1 | $15.00 | $10.00 | $25.00 |

The same job ranges from a quarter to $25. If a small model's summaries pass your quality check, the choice is easy.

> **Try it:** Take one real task you do with AI, estimate its input and output tokens (or paste a sample into a token counter), and fill in the worked-example table for the cheapest and the most expensive model you would consider.

## How do you keep the bill down?

**1. Start with the smallest model that passes your test.** Write ten examples with known good answers, try the cheap model first, and move up only if it fails.

**2. Cache what repeats.** If every request starts with the same long instructions or document, prompt caching stores it. On Claude, a cache hit costs 10% of the normal input price; on OpenAI, cached input for GPT-6.1 Sol is $0.10 per million instead of $2.

**3. Batch what can wait.** Both companies charge half price for work submitted in bulk and returned later, such as overnight summaries. On Claude, batch and caching discounts combine.

**4. Watch the long-context premium.** Very long prompts cost more per token. OpenAI charges GPT-6.1 Sol $4 instead of $2 per million input tokens for long-context requests, and Claude Haiku 5.5 jumps from $0.10 to $0.50 once a prompt passes 100,000 tokens. Send only the parts of a document the task needs.

**5. Skip speed tiers you do not need.** OpenAI's Fast option doubles the price and Ultrafast multiplies it by six. Tools also add charges: Claude's web search costs $10 per 1,000 searches plus the tokens of the results.

**6. Limit the output.** Output is the expensive half. Ask for "three bullet points" rather than "a summary", and set a maximum output length in your code.

**7. Set a spending cap.** API consoles let you set monthly spend limits. Set one before you run anything in a loop.

For most people, the sensible split is a $20 subscription for daily chat and a small API budget for experiments. Pay for a bigger plan only when you keep hitting the limits, not because a bigger model exists.

## Sources

- Pricing: https://platform.claude.com/docs/en/about-claude/pricing
- Pricing: https://developers.openai.com/api/docs/pricing
- Plans & Pricing: https://claude.com/pricing