LLM Pricing Models: Per Token, Per Request, or Subscription?

LLM pricing usually falls into three patterns: usage charges based on tokens, charges based on requests or tasks, and recurring subscriptions with a defined level of access. Token billing is common for developer APIs, while subscriptions are common in chat products; request-based pricing appears when a service charges for a whole operation rather than each token.

What the pricing models mean

A token is a small piece of text processed by a language model. It may be a whole short word, part of a longer word, punctuation, or a space-related unit; tokenization varies by model and language. In a typical API, the bill separates input tokens—the prompt, instructions, and supplied context—from output tokens generated in response. Some services also distinguish cached input, tool use, images, audio, or other modalities.

Per-request pricing charges a set amount for an operation, or uses a fixed unit such as a task, document, or image. It can be easier to predict when each request represents a similar amount of work. But “per request” does not necessarily mean every request costs the same: a service may define tiers, limits, or different prices for different operation types.

A subscription charges a recurring fee for access to a product or a bundle of features. It may include a usage allowance, limits on messages or speed, access to certain models, or fair-use restrictions. A subscription is not automatically unlimited, and its allowance may not translate into a fixed number of tokens.

How token billing works

For a basic text API estimate, multiply the input and output token counts by their respective rates, then add the results. If a hypothetical service charged $2 per million input tokens and $8 per million output tokens, a call using 10,000 input tokens and generating 2,000 output tokens would cost:

• Input: 10,000 ÷ 1,000,000 × $2 = $0.02
• Output: 2,000 ÷ 1,000,000 × $8 = $0.016
• Total: $0.036, before any other applicable charges

These rates are illustrative, not current prices for a particular provider. Actual rates differ by model and service, and providers can change them. A real estimate should use the provider’s current pricing page and account for any separate charges for tools, storage, data processing, or non-text inputs.

Input is not limited to what a person types into a chat box. Repeated system instructions, conversation history, retrieved documents, and tool results can all contribute to input usage. Long answers increase output usage. Some services offer discounts for cached or batch processing, subject to their specific rules.

Request charges and subscriptions compared

Request-based charges can suit a workflow where each operation has a clear boundary—for example, processing one uploaded document or generating one image. The key question is what counts as one billable request. A single user action might trigger multiple model calls, retries, or tool operations, and a service may charge for each underlying operation rather than the action as a whole.

Subscriptions are often simpler for individuals who use a product regularly and value an easy-to-understand recurring bill. They may bundle a chat interface, file handling, or other features. However, access rules can vary, and a subscription for a consumer-facing app may not include API credits. A user may need a separate account or billing arrangement to build software using an API.

Token billing gives developers a closer link between model usage and cost. It can be economical for light or uneven workloads, but monthly totals can fluctuate. A subscription makes recurring spending more predictable up to its limits, but may be poor value if it goes mostly unused or if extra usage requires another plan or payment.

How to estimate your real cost

Start with the work, not just the headline rate. Measure or estimate how often the system runs, how much context each call includes, and how long responses tend to be. Then check which model and features the task actually requires.

1. Count complete interactions. Include instructions, conversation history, retrieved content, and tool output—not only the latest question.
2. Separate input from output. They may have different rates. For a chat product, check whether usage is metered in tokens at all.
3. Check non-text charges. Image, audio, video, search, and other tools may use separate units or pricing rules.
4. Allow for retries and peaks. Failed calls may or may not be billed; verify the service’s policy. Production systems can also have bursts that differ from an average day.
5. Compare the same workload. Test realistic examples with each candidate model or plan, including response quality and latency needs.

For an API, a simple monthly estimate is the expected number of calls multiplied by the average cost per call. This is only a planning estimate: prompt length, output length, model choice, caching, and feature use can vary. Usage dashboards and billing alerts, where available, help reveal surprises after deployment.

Choosing a model for your use case

For occasional personal questions, compare the subscription’s included access and restrictions with how often you will use it. If you need to integrate a model into an application, compare API pricing and terms separately; consumer subscriptions usually should not be assumed to cover API usage.

For variable workloads, token-based billing can avoid paying a flat monthly fee for idle capacity. For repeated, predictable operations, a per-request plan may simplify budgeting if the billable unit matches the work. For steady, frequent use, a subscription or committed plan may be convenient, but only after checking its limits and what happens when they are reached.

Also compare capability, not just cost. A lower-priced model may need more retries, longer prompts, or human review to meet a task’s quality needs. Conversely, using a more capable model for every simple operation may add expense without a useful improvement. A common approach is to test a smaller model for routine tasks and reserve a more capable one for cases that need it.

Common billing surprises

• Confusing tokens with words: there is no universal word-to-token conversion. Text, language, and tokenization affect the count.
• Forgetting conversation history: a continuing chat may resend earlier messages as context, increasing input usage.
• Assuming a flat plan is unlimited: subscriptions can impose usage caps, throttling, or feature-specific restrictions.
• Mixing product and API billing: an app subscription and developer API account can be separate products.
• Ignoring add-ons: tools and media processing may be charged separately from text tokens.

Before relying on a quoted rate, read the billing unit, included allowance, overage rules, and cancellation or renewal terms on the service’s own pricing and plan pages. Pricing can change, so dated comparisons may no longer be reliable.

Frequently Asked Questions

Is per-token pricing cheaper than a subscription?
It depends on how much you use the service and what the subscription includes. Light or occasional API use may cost less when metered, while frequent use may make a subscription attractive if it covers the needed features and usage.

How many tokens are in a word?
There is no exact universal ratio. Token counts depend on the text and the model’s tokenizer, so use the provider’s tokenizer or usage data for a useful estimate rather than treating words as tokens.

Does a chat subscription include API usage?
Not necessarily. Consumer chat access and developer API billing are often separate. Check the specific plan’s terms instead of assuming one pays for the other.

Are input and output tokens priced the same?
Not always. Many APIs list separate rates, and some distinguish cached input or other processing types. Use the exact categories shown in the current pricing information.

What is the easiest way to keep API costs down?
Track usage, avoid sending unnecessary history or documents, limit output to what the task needs, and choose a model that meets the quality requirement without excess capacity. Check the provider’s rules before relying on caching or batch discounts.

Related

• AI Job Displacement: What It Means and How to Prepare
• Best HIIT YouTubers for Beginners to Burn Calories
• Micro1’s $1 Billion Data-Buying Plan Explained