Token-Based Pricing For AI Products: Why It Looks Simple But Can Hurt Margins

In our previous article, Hybrid Pricing for AI Agents: The Practical Model Between Subscription and Usage, we showed how a subscription can provide predictability while a variable charge captures additional usage.

The next question is what that usage should be measured in.

Tokens appear to be the simplest answer. They are measurable, familiar to technical teams and directly connected to foundation model costs.

But a token is an infrastructure unit, not necessarily a product-value unit.

One customer-facing AI agent action can trigger several internal model calls, tool executions and retrieval steps. The customer sees one completed task. The vendor pays for the entire chain.

Token Pricing Works at the Infrastructure Layer

Token-based pricing is not inherently wrong.

It works well for foundation-model providers because they sell model inference. The billing unit and the underlying cost unit are closely connected.

As of July 10, 2026:

• OpenAI lists GPT-5.6 Terra at $2.50 per million input tokens and $15 per million output tokens.
• Anthropic lists Claude Sonnet 5 at an introductory $2 per million input tokens and $10 per million output tokens through August 31, 2026. Standard pricing then moves to $3 and $15.
• Google's Gemini API pricing also separates input, output, cached-token and storage charges.

For a developer buying direct API access, this structure makes sense. For an application-layer AI product, it may not.

The customer is rarely buying tokens. They are buying a resolved ticket, completed workflow, generated report or processed document.

One Agent Action Can Trigger Several Model Calls

A customer may submit one request, but an AI agent may need to:

1. Interpret the task
2. Plan the workflow
3. Retrieve information
4. Select and call a tool
5. Evaluate the result
6. Generate the final answer

OpenAI describes the agent loop as an iterative process in which the model runs inference, requests tools, receives their results and may query the model again. Its developer guidance also notes that agents can call tools and execute functions over multiple turns when needed.

Anthropic similarly describes agentic systems as language models augmented with retrieval, tools and memory.

The number of calls varies by product and task. That variability is exactly why pricing from the visible prompt and response alone can be dangerous.

Worked Example: One Conversation, Six Model Calls

Consider an AI support product that charges $0.10 per conversation.

The price assumes each conversation contains:

• 2,000 input tokens
• 500 output tokens
• 2,500 visible tokens in total

At 10,000 conversations per month:

10,000 × $0.10 = $1,000 monthly revenue

Now assume, for this illustrative workflow, that each resolved conversation triggers six model calls.

Six is not an industry average. It represents a multi-step agent loop involving planning, retrieval, tool use, evaluation and final response generation.

If every call uses 2,000 input tokens and 500 output tokens, total monthly consumption becomes:

Input tokens: 10,000 conversations × 6 calls × 2,000 tokens = 120 million input tokens

Output tokens: 10,000 conversations × 6 calls × 500 tokens = 30 million output tokens

Using GPT-5.6 Terra's published price of $2.50 per million input tokens and $15 per million output tokens:

Input cost: 120 million × $2.50 per million = $300

Output cost: 30 million × $15 per million = $450

Total model cost: $750

The product earns $1,000 and spends $750 on inference alone.

If the company had calculated cost using only one visible model call per conversation, it would have estimated a cost of $125 and a gross margin of 87.5%. Internal call fan-out reduces the expected margin from 87.5% to 25%.

That still excludes search, embeddings, storage, tools, APIs, orchestration and human review.

Three Ways Token Pricing Erodes Margin

1. Visible Usage Is Not Total Usage

Customers see the prompt and final answer. The product may also consume tokens for planning, retrieval, tool selection, retries, verification and communication between agents.

If that internal consumption is not captured in the price, the vendor absorbs it.

2. Equal Token Volumes Can Create Different Costs

Input and output tokens are priced differently. Output tokens can cost substantially more, as the current OpenAI, Anthropic and Google pricing tables demonstrate.

Agents may also route tasks across different models.

Two customers can consume the same total number of tokens while creating different costs because one generates longer outputs, uses larger contexts or triggers more expensive models.

A single blended token rate can hide that difference.

3. Tokens Do Not Capture the Whole Workflow

An AI agent may also generate costs through:

• Tool calls
• Search
• Embeddings
• Vector databases
• Storage
• Third-party APIs
• Code execution
• Human review

A token-only price can look profitable while the complete workflow loses money.

Token Based Pricing vs Usage Based Pricing

Token-based pricing is one form of usage-based pricing, but they are not the same.

Token-based pricing charges for model input and output.

Usage-based pricing can charge for any measurable product activity, such as:

• Agent actions
• Workflow runs
• Documents processed
• Resolutions
• API calls
• Completed outcomes

Tokens are close to infrastructure cost.

Actions, workflows and outcomes are often closer to the value customers understand.

This continues the shift examined in Why Seat-Based SaaS Pricing Fails for AI Agents: once software performs the work, neither a human seat nor a raw infrastructure meter necessarily captures the complete value delivered.

Intercom, for example, charges $0.99 for a successful Fin outcome, rather than exposing raw token usage to the customer. The customer pays when the AI agent completes useful work, even when that outcome requires multiple internal actions.

The Better Model: Track Tokens, Price the Work

AI companies should not stop tracking the tokens. They should stop treating tokens as the entire customer-facing pricing model.

A stronger token to value model has three steps.

1. Track the True Cost

Capture every input, output, cached and reasoning token across all internal model calls.

Add tool, API, storage, search and infrastructure costs to calculate the real cost of each customer, agent and workflow.

2. Price a Unit the Customer Understands

Translate that internal cost into a product unit such as:

• One workflow
• One resolution
• One processed document
• One agent action
• One credit
• One completed analysis

The customer pays for what the product delivers, not the architecture behind it.

3. Apply Cost-Plus Margin Guardrails

Use the full delivery cost to establish the minimum sustainable price.

Then consider the business value of the completed work.

The final price must sit above cost while remaining aligned with what the result is worth to the customer.

This is an agent-native approach to AI product pricing because it connects both sides of the transaction:

What the agent costs to run and what its work is worth.

How Revinci Connects Token Usage To Margin

At Revinci, we treat token usage tracking as one part of the revenue model, not the final pricing decision.

Revinci Bill supports tokens as financial and billing units while connecting them with agent actions, pricing, wallets and margin intelligence.

Teams can track tokens internally and use that data to support token-based, usage-based, hybrid or outcome pricing without exposing every infrastructure event to the customer.

Our SmartCost and SmartMargin intelligence connects agent revenue with token, compute, storage and API costs. Teams can then assess cost-to-serve and margin by customer, agent or workflow.

A usage-based billing platform for AI should not answer only:

How many tokens were consumed?

It should also show:

• What created those tokens
• Which customer and agent consumed them
• What the complete workflow cost
• What the customer paid
• What margin remained

Accurate token metering is useful. Accurate token metering without margin visibility is not enough.

Know Every Token. Do Not Sell Every Token

Token pricing belongs naturally at the foundation-model layer because model providers sell inference. One visible agent action can trigger several model calls with different input, output and model costs. It may also rely on tools and infrastructure that token consumption does not capture.

Tokens should remain central to AI usage tracking and AI cost tracking.

But the customer-facing price should reflect the work the product delivers.

The goal is not to bill every token burned.

It is to know every token burned, understand its true cost and price the resulting work at a margin the business can sustain.

Frequently Asked Questions