Buy That Same Token: Put a Dollar Figure on Every AI Task You Run

Written by John Steele | Jun 12, 2026 4:52:30 PM

Ask a founder what their last heavy afternoon in Claude or ChatGPT cost them and you'll get a shrug.

The reassuring part: nobody is getting this right.

Uber blew through its entire 2026 AI coding budget in four months.
A four-person startup ran up $113,000 in a single month.
A Priceline team watched a routine Cursor renewal come back four to five times higher.

Enterprise is responding. Citi posted a VP-level FinOps role whose whole job is controlling token costs, a role that didn't exist two years ago. And Salesforce is building a system that routes each task to the cheapest model that can handle it.

Fortunately, the fix for startups is lighter, and you can build it in an afternoon. What you need is exactly what the big guys need. But with fewer tasks, you have more ability to plan, track and execute. What you need is an AI task log that surfaces:

How many tokens each task used
How long the task took
What model or models were used
The route the AI took to get to the goal
How much the task cost

Two pieces make it work. First you have to run all tasks in Claude Code because it is the only Claude feature that delivers an exact, per-task token count. The chat apps don't, and the models can't estimate. And second, you have to choose the model you plan to use in advance because cost swings hard between them: Opus runs about five times the price of Haiku.

Are you ready to spend your tokens smarter than ever. Let's get it.

	We're building two things: A templated task structure to capture data from the human user, and a script to capture data directly from the AI model. Both will be logged just like any other task you mark complete.

Normal software doesn't cost more to use after an update. AI does.

A version bump can quietly raise what the same work costs, with no notice and no new contract. Anthropic's Opus 4.7 shipped with a new tokenizer that can produce up to 35% more tokens than Opus 4.6 for the same input.

Different models do different things well. And prices may vary.

Advanced reasoning is what the labs lead with at every launch, from every stage and in every interview. But reasoning is expensive, and if you're just brainstorming titles for your next email, you could be burning money.

You can't act without data.

The AI models don't make it easy. A model asked to count its own tokens just guesses. And chat apps have proven to be very bad at budgeting. So you need to have access to your own tasks, costs and data usage.

› Every AI task ends up with a record it never had: the model used, time to completion, tokens spent, the plan Claude proposed and you approved, and how it turned out. That record sits in ClickUp for reporting.
› The log answers the burning questions like which model is right for which job, what your steady, repeatable processes cost at baseline, whether a task needed advanced reasoning at all, or you paid for it out of habit, how long AI takes and would a human be faster at the same task.
› The notification layer (available within ClickUp natively) keeps the approval process moving.

Prerequisites:

Claude Code installed

Claude Code is the non-negotiable piece, for the reason above: it writes the per-task token log this whole thing reads from. If it isn't set up on your machine, let's talk--I can get you sorted.

A ClickUp account and an API token

ClickUp is free to test this on, and $7 a month (the Unlimited plan) to run it at scale, since the custom fields that hold the log are capped on the free tier. You can try a different task management platform, but their custom field terms are less favorable.

› Build the ClickUp template: Create a List called AI Task Log. Give it four statuses: Ready, In Progress, Pending Approval, Done.

Add these custom fields:

Brief (long text)
Model (dropdown, Opus, Sonnet, Haiku, Fable)
Approved Plan (long text): the plan Claude proposes in Plan Mode, pasted in when approved.
Input Tokens(number)
Output Tokens (number)
Equivalent API Cost (currency)
Actual Cost (currency)
Outcome (long text)

The four statuses timestamp themselves as the task moves, so you get time-to-completion without adding a field for it. Save the List as a Template. Every ClickUp task already has a unique ID, and that ID is what ties the work back to the right row. Record both cost fields: equivalent cost is API rates, actual cost is what you paid (zero on a subscription).
› Get your token and field IDs: In ClickUp, go to Settings → Apps and generate an API token. It starts with pk_. Keep it private. You also need the ID of each custom field. Don't dig for these by hand. Open Claude Code and paste this, with your token and List ID filled in (your List ID is in the URL when you open the List):

Run this and show me the field names and IDs:
url -s -H "Authorization: pk_YOUR_TOKEN" "https://api.clickup.com/api/v2/list/YOUR_LIST_ID/field" | jq -r '.fields[] | .name + ": " + .id'

› Let Claude Code create the approve command: Once you create this, every time you type /approve and hit enter, Claude Code will read the session's tokens and the model from the log, fetch the current rates from Anthropic, compute the cost, post the numbers to your ClickUp task, and close the task. Copy the whole prompt from this Doc: Funnel Vision 22 /approve command. Fill in your ClickUp API token (replace pk_YOUR_TOKEN) and the four field IDs from Step 2, and paste it all into Claude Code.

Start a fresh Claude Code session for each task. The /approve prompt totals the most recent session log, so one task per session keeps each count its own. The first time you run /approve, Claude will ask permission to fetch anthropic.com/pricing. Allow it and it won't ask again.

› Add the Slack nudge: CSlack tells you when to come back and approve. You approve in Claude Code, not in Slack. Slack is just the ping.

In ClickUp, go to Settings → Integrations → Slack and connect your workspace.
Open your AI Task Log List, click Automations, then Add Automation.
Set: When status changes to Pending Approval → Post to Slack in your channel, with the message "Plan ready for review:" plus the task name variable from ClickUp's picker.
Add a second one: When status changes to Done → Post to Slack, message "Task closed, cost logged."

› Run a task: To run a task, start a fresh Claude Code session, enter plan mode by typing /plan (or pressing Shift+Tab twice), and paste the brief behind a one-line wrapper:

Task ABC-123. Here is the brief: [paste the Brief field from ClickUp]. Plan this first and show me the plan. Don't change anything until I approve it.

Plan mode keeps Claude read-only until you accept, so you see the approach before any work runs.

When the AI subscription comes up for renewal, you bring task-level evidence instead of a gut feeling. If you are a founder, you can begin accounting for the success (or failure) of AI intervention. Some tasks will be faster and more efficient. Some will not. But at least you'll know, even as the models change.

Full disclosure: this runs on Claude Code, and getting that set up is something I do for clients. If you want a hand standing it up, or you want the automated version, book a 30-minute intro call.

JPMorgan, Salesforce, Google, Microsoft, and IBM are funding a new industry standards body for AI token cost management, debuting at FinOps X this month. The group aims to bring the FinOps cost-discipline model to AI token spend. Learn More....

Uber has imposed a $1,500 monthly cap on individual engineer AI tool usage. The company has created dashboards engineers can follow and has set competitions to reduce token usage. The company hopes the move will help after blowing through their annual AI budget in 4 months. Learn More...

DeepSeek's 75% V4-Pro discount became permanent on May 31, with cached input down to a tenth of the old rate. The new rates could entice enterprises like Disney and Visa that run tens of thousands of workloads per day. Learn More...

View full post