AI pricing can be confusing. One product charges a monthly fee per user, another bills per million tokens, a third sells credits that vanish at different speeds depending on what you do. Behind the variety are a few simple models. Understanding them helps you choose plans and estimate costs before they surprise you.

Why AI pricing is different

Traditional software costs almost nothing to run per extra user. AI is different: every request consumes compute, and longer or harder requests consume more. Vendors design pricing to cover that variable cost while keeping it simple for customers.

Model 1: subscriptions (per seat)

Consumer and business assistants usually charge a flat monthly fee per user, often with tiers:

  • Free: limited usage and features.
  • Individual paid plans: higher limits, stronger models, extra features.
  • Team/business plans: admin controls, data protections, shared workspaces.
  • Enterprise: negotiated pricing, security reviews, compliance features.

“Unlimited” plans almost always have fair-use limits, especially for the most expensive models or features.

Model 2: pay per token (APIs)

Developer APIs charge by tokens, the chunks of text models read and write. Prices are usually quoted per million tokens, with separate rates for:

  • Input tokens: everything you send — instructions, documents, conversation history.
  • Output tokens: everything the model writes. Output is typically several times more expensive than input.

Additional rules often apply:

  • Cached input — repeated content such as a long system prompt — may be billed at a large discount.
  • Batch processing — requests that can wait hours — is often discounted.
  • Reasoning models may bill for internal “thinking” tokens.
  • Images, audio and video are converted to token equivalents or priced separately.

Model 3: credits

Many creative tools — image, video and audio generators — sell credits. Each action costs a number of credits that depends on resolution, duration or model quality. Credits simplify pricing for users but make costs harder to compare across tools.

Estimating API costs: a worked example

Suppose a support assistant handles 10,000 conversations a month. Each conversation sends about 3,000 input tokens (instructions plus history) and receives 500 output tokens.

  • Input: 10,000 × 3,000 = 30 million tokens
  • Output: 10,000 × 500 = 5 million tokens

Multiply each by the provider’s price per million. Then add a margin of 20–30% for retries, longer conversations and growth. Comparing two models this way often reveals that a smaller model is many times cheaper for acceptable quality.

Hidden costs

  • Conversation history: each turn resends earlier messages, so long chats cost more per message over time.
  • Retrieval context: stuffing many documents into every prompt raises input costs quickly.
  • Failed attempts and retries.
  • Seats for occasional users in subscription plans.

Tips

  1. Start with usage estimates from a small pilot, not guesses.
  2. Set budget alerts and spending limits in API dashboards.
  3. Use the cheapest model that passes your quality tests.
  4. Revisit prices regularly: they have generally fallen for a given level of capability.

Common mistakes to avoid

  • Estimating costs from a single test prompt. Real usage includes longer conversations, retries and edge cases. Measure a pilot with real users.
  • Forgetting that history is resent. In chat applications, every new message usually includes the earlier conversation, so costs per message grow over a long chat.
  • Buying seats for occasional users. Many people use AI tools only now and then. Check usage reports and adjust licenses.
  • Ignoring currency, tax and minimums. Enterprise contracts can include minimum commitments, and prices may be quoted in dollars before tax.
  • Not setting spending limits. A bug or loop in an automated system can generate large bills quickly. Use budget alerts and hard limits.

Key takeaways

  • AI pricing reflects compute: every request has a real running cost.
  • Assistants usually charge per seat; APIs charge per million input and output tokens; creative tools often use credits.
  • Output tokens cost more than input; caching and batch processing can cut costs significantly.
  • Estimate API costs from real token counts in a pilot, then add a safety margin.
  • Long conversations and large retrieved contexts are the most common hidden cost drivers.

Frequently asked questions

How many words is a million tokens?

In English, roughly 750,000 words, since a token averages about three-quarters of a word. Other languages and code often use more tokens for the same content.

Why are output tokens more expensive?

Generating text happens one token at a time and is computationally heavier per token than reading input, which models process in parallel.

Are subscription plans cheaper than the API?

For an individual using a chat interface heavily, a subscription is usually cheaper. For automated, high-volume or embedded use, the API is the right tool, and costs scale with usage.

Do AI prices go down over time?

Historically, the price for a given level of capability has fallen substantially as models and hardware became more efficient, although the newest, most capable models are usually priced at a premium.