How do I find my real token counts?

Every API response reports them. Anthropic returns input_tokens, output_tokens, cache_read_input_tokens and cache_creation_input_tokens in usage; OpenAI and Gemini return the same counts under their own names. Log them for a week and use the averages per call here.

What counts as a task?

One thing a user asks your product to do: an agent run, a generated report, a chat answer. A task can make several model calls. An agent that plans, uses tools and answers is often 10 to 30 calls.

Why does turning caching on make a task cost more at first?

Anthropic and OpenAI charge 1.25x the input price to write a prompt to the cache. Caching pays off once most of each call is read back at the cache price, which is why agents that resend the same context every turn save the most.

Why plan for the heavy user and not the average?

With AI, every action has a price, so a few users can cost more than everyone else combined. A plan is safe when the user who maxes it out still leaves a margin. That is what the included usage and overage numbers on this page are built for.

How current are the prices?

They come from the providers' own pricing pages and are checked once a month. Last check: 2026-10-05. They are standard rates for prompts under 200k tokens, without batch or negotiated discounts.

Build with AI · AI Feature MarginPrices checked Oct 5, 2026

Free tool

AI Feature Margin Calculator

Put in how your AI feature calls the model and what you charge for it. You get the cost of every task, the margin on your average and your heaviest user, and the usage cap and overage price that keep the plan profitable, on 10 current models from Anthropic, OpenAI and Google.

1. One task

Model

Model calls per taskInput tokens per callOutput tokens per call

Input read from the prompt cache0%0% means caching is off. Agents that resend the same context often reach 80–90%.

2. Your plan and your users

Plan price per month$Other cost per user per monthHosting, email, support$Tasks a month, typical userTasks a month, heaviest user

Cost per task

$1.90

on Claude Sonnet 5.5

Input: $1.60
Cache reads: $0.00
Output: $0.30

Typical user

$14.30

72% margin · AI cost $5.70

Heaviest user

−$398.00

−1990% margin · AI cost $418.00

Break-even

10 tasks

a month, before a user costs you more than they pay

Include in the plan

7 tasks

keeps 30% margin on a user who uses them all

Overage price

$3.17 per task

past the cap, keeps 40% margin on every extra task

Your heaviest user costs you $398.00 a month more than they pay. Try turning on caching, moving to a cheaper model below, or capping the plan at 7 tasks with overage after that. The AI Feature Margin playbook shows how to do each one.

The same task on every model

Your workload above, priced on each model, cheapest first. Quality isn't on this table: test a cheaper model on your own tasks before you move to it.

Model	Per task	Heaviest user / month	Margin on them
OpenAI	$0.0950	$20.90	−4%
Google	$0.32	$69.30	−246%
Google	$0.71	$156.75	−684%
Anthropic	$0.95	$209.00	−945%
Anthropic	$1.90	$418.00	−1990%
OpenAI	$1.90	$418.00	−1990%
Google	$1.96	$431.20	−2056%
Anthropic	$3.80	$836.00	−4080%
Anthropic	$9.50	$2,090.00	−10350%
OpenAI	$9.50	$2,090.00	−10350%

Standard prices per million tokens for prompts under 200k tokens, checked Oct 5, 2026 on the pricing pages of Anthropic, OpenAI and Google. No batch, regional or negotiated discounts.

How the numbers are calculated

Cost per task = calls × (input tokens × input price + output tokens × output price), per million tokens.
With caching, the cached share of the input is billed at the cache-read price and the rest at the cache-write price (1.25x input on Anthropic and OpenAI). Gemini's hourly cache storage fee is left out.
Margin per user = plan price − other cost − tasks × cost per task.
Include in the plan is the most tasks a user can run while the plan still keeps 30% of its price. Overage is priced so each extra task keeps 40%.
The page opens on the example agent from the AI Feature Margin playbook: 20 calls, 40,000 tokens in and 1,500 out per call, on a $20 plan.

Get the numbers from your own product

The playbook has the log table and the SQL that find your real tokens per call and your heaviest user, plus the three fixes in order: caching, model routing, included usage.

Read the playbook Learn AI unit economics in the Advanced AI Product Manager masterclass

Questions

How do I find my real token counts?: Every API response reports them. Anthropic returns input_tokens, output_tokens, cache_read_input_tokens and cache_creation_input_tokens in usage; OpenAI and Gemini return the same counts under their own names. Log them for a week and use the averages per call here.
What counts as a task?: One thing a user asks your product to do: an agent run, a generated report, a chat answer. A task can make several model calls. An agent that plans, uses tools and answers is often 10 to 30 calls.
Why does turning caching on make a task cost more at first?: Anthropic and OpenAI charge 1.25x the input price to write a prompt to the cache. Caching pays off once most of each call is read back at the cache price, which is why agents that resend the same context every turn save the most.
Why plan for the heavy user and not the average?: With AI, every action has a price, so a few users can cost more than everyone else combined. A plan is safe when the user who maxes it out still leaves a margin. That is what the included usage and overage numbers on this page are built for.
How current are the prices?: They come from the providers' own pricing pages and are checked once a month. Last check: 2026-10-05. They are standard rates for prompts under 200k tokens, without batch or negotiated discounts.