The three ways you pay for AI
- A flat monthly subscription. Like Netflix. You pay $20 (or $100, or $200) a month and get access to the chat product with some usage limits.
- Pay-per-token through the API. You only pay for what you use, billed in fractions of a cent per token. Used by developers and businesses building on top of AI.
- Free plans with caps. Limited messages per day, often on a smaller model. Good for testing, not for daily work.
Subscriptions, the everyday option
If you're a regular user, this is what you want. Both ChatGPT and Claude charge similar amounts.
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Limited messages, basic models, no advanced features. Fine for trying things out. |
| Pro / Plus | $20/month | Much higher limits, access to the best models, file uploads, web search, image generation, custom instructions. |
| Max / Pro Plus | $100–$200/month | The most generous limits, priority access, longer context, advanced reasoning modes. |
| Team | $25–$30/user/month | Everything in Pro plus shared workspaces and admin controls. Best for 2+ person teams. |
| Enterprise | Custom | For larger orgs. Adds security controls, audit logs, single sign-on, and contractual data protections. |
How "message limits" actually work
Pro plans usually advertise something like "400 to 2,000 messages every 5 hours." This is not a per-message cap. It's a moving window. You can burn through a lot if every message is short, or hit the cap faster if every message includes a 50-page document.
The real limit is total compute. Big documents, long reasoning, and use of the most capable model eat through your allotment faster. A normal day of writing emails and asking questions will not get you anywhere close to the cap.
You're either using the most expensive model for tasks that don't need it, or you've left a long file in the chat that gets re-processed on every message. Start a new chat or switch to a smaller model.
The API — pay-per-use option
The API is for developers and businesses building AI into their own tools. You don't see the chat interface. You pay per token and only for what you actually send and receive. No monthly minimum.
API makes sense when:
- You're building automation that runs without a human watching it
- You want AI inside your own software, not in someone else's chat window
- Your usage is unpredictable — sometimes a lot, sometimes nothing
- You need higher rate limits than a chat subscription allows
API does not make sense when you're a single person who just wants to use AI to write and think. The subscription is almost always cheaper for that.
Hidden costs to watch for
- Image generation. Counts as separate usage on most plans. Heavy image work can blow through limits fast.
- Web search. Some plans charge for it. Others include it. Check before you assume.
- File uploads. A 100-page PDF can be 50,000+ tokens. Three of those in one chat is a real bite.
- The "thinking" or "reasoning" mode. Models like Claude Opus 4.7 and GPT-5.5 Pro burn way more tokens when they think step-by-step. Only turn it on when you actually need deep reasoning.
- Background agents. Tools that run in the background can keep working and racking up tokens after you stop watching.
How to pick the right plan
- Just exploring? Free. Try ChatGPT and Claude side by side for a week before paying.
- Daily user, one person? $20/month Pro plan from whichever AI you prefer.
- Power user or running a business on it? Max or Pro Plus ($100–$200/month). The bigger allotment matters once you're using AI for hours a day.
- Small team, 2–10 people? Team plan. Shared workspaces and centralized billing matter more than you'd guess.
- Law firm, regulated industry, or any client data involved? Enterprise. The data protection contracts and audit logs are non-negotiable for compliance.
- Building software? API. Start with the cheaper models (Haiku, GPT-5.5 nano) and only upgrade where you need to.
The advanced take: optimizing a real bill
For a business running AI across multiple workflows, the cost saving tactics are:
- Match the model to the task. Use Haiku for sorting, Sonnet for writing, Opus only when reasoning really matters. This alone can cut bills 60–80%.
- Use prompt caching. If your system prompt is reused, you pay 90% less on the cached portion after the first call.
- Batch jobs that aren't urgent. Both Anthropic and OpenAI offer batch processing at 50% off if you can wait a few hours.
- Set spending limits. The API lets you cap monthly spend so a runaway script doesn't surprise you.
- Track per-user usage on Team plans. One person doing transcription work can easily use 10x what a writer uses.