For businesses that run their own AI agents

Optimize your AI costs by what actually gets done.

See the real cost of every successful job, spot the agents that waste money, and fix them first. Runs on your computer. Never keeps your data.

Free to use · Windows coming soon

See the cost. Spot the waste. Fix it first.

The Results screen · example data

Results

September 2026 · figures below for Support replies · OpenAI and Claude
Cost per finished job$0.39AI bill, retries included · indicative
Spent on jobs that didn't finish38.2%$1,160 of $3,040
Jobs finished7,800of 9,800 with an outcome
Compared with a person6%of your estimate for a person
Example agents and what each finished job cost
AgentJobs doneCost / finished jobOn failuresIn plain words
Support replies
Customer support
7,800$0.3938%A bad job costs 3.2× a typical one. A per-job limit would catch the runaways.
Lead research
Sales leads
1,240$0.8422%Steady from job to job.
Invoice entry
Data entry
3,410$0.0661%Most of the money goes on jobs that don't finish.
Blog drafts
Content
0——Spending, nothing finished $46.20 spent and no job has finished yet.

Example data. All figures are indicative. "Compared with a person" uses your own estimate of the minutes a person takes and what their hour costs.

The problem

Your AI bill counts attempts, not finished jobs.

49%

got a higher AI bill than they expected

Of those paying for AI, in a Q3 2026 survey of 1,150 decision-makers at businesses with 5 to 499 staff. 57% in the US.

You pay for every try

Each time an agent runs, you pay, whether the job got done or not. The bill can't tell you which runs actually worked.

Cost of each run, same jobillustration

The same job, very different prices

One run of a job can be cheap and the next can cost many times more. An average hides the expensive ones.

Some costs never show up

Failed and retried attempts are paid for, then forgotten. So is the time your staff spend checking and fixing AI work. Neither is a line on the bill.

55%

spend 3 or more hours a week checking AI output

Time that never appears on an AI bill, but is part of what each finished job costs.

Source: Pax8 SMB AI Pulse, Q3 2026, a survey commissioned by Pax8, which sells to small businesses. Press release, 12 Oct 2026 · TNW report

How it works

Three steps to the real cost of each job.

Step 1

Connect OpenAI or Claude

Add an admin key, or drop in the usage file you can download from OpenAI or Claude, and CostGrid imports your last 90 days. The key goes into your Mac's Keychain, and if you give it an expiry date, CostGrid stops using it then. Then say which agent uses each key.

Step 2

See which jobs got done

If your agents send requests through CostGrid (a small change for whoever set them up), CostGrid estimates which jobs finished from how each agent behaved. Their software can also report it exactly. Otherwise, enter how many jobs finished each month.

Step 3

See the real cost per job

Failed and retried attempts are charged to the jobs that worked. Each agent gets one plain sentence, such as steady, mostly failing, unpredictable, too early to say, or spending with nothing finished. Add how long a person takes and it compares the two. All figures are indicative.

Fix it first

Spend less on the same work.

Send your agents' requests through CostGrid and it can cut what each job costs, not just report it. Needs a small change by whoever set your agents up.

Move simple work to a cheaper model

A rule sends a model's calls to a cheaper one from the same provider. CostGrid can propose it from your own traffic, and it runs as a dry run first, showing what it would have changed and saved, before anything switches.

Switch instead of stopping

When an agent goes over its budget, CostGrid can move it to a cheaper model rather than refuse its calls, so the work carries on at a lower rate.

Check that it actually saved

A cheaper model can cost more if it thinks longer or takes more steps. CostGrid compares the cost per call and per whole run before and after on your own traffic, and shows a switch that cost more as a loss. Estimates are labelled as estimates.

Example result

One support ticket, all‑in.

The AI bill is only part of what a finished job costs. Here is one support ticket, with everything added up. Illustrative, from the LCSO method's worked example.

Cost of one finished support ticket

AI bill only
$0.39
All-in cost with AI
$1.52
A person doing it · estimate
~$6.75
One shared scale, $0 to $8. The person figure is an estimate: 9 minutes at $45 an hour, from a time study in the worked example. Illustrative numbers, not a customer's.

Where the $1.52 comes from

The attempt that worked
$0.24
Failed and retried attempts
$0.15
AI bill
$0.39
Staff checking and fixing · estimate
$0.62
Fixed costs, shared across the month
$0.51
Real cost per finished ticket
$1.52

Illustrative, from the LCSO method's worked example. Staff time is an estimate from a time study. Fixed costs cover tools and upkeep, shared out over the month's finished tickets. CostGrid's Results screen shows the AI-bill part for each agent; the all-in figure is the method's full picture.

Pricing

Start free. Add control when it's ready.

View

Free

See where the money goes.

  • Costs per agent
  • Cost per successful job
  • Share of spend on failed attempts
  • How much the cost of the same job varies
  • Comparison with your estimate for a person
  • Warnings about risky key setups
  • Move work to a cheaper model (for traffic through CostGrid)

Failed-attempt share and cost spread need the agent's requests to go through CostGrid; for other agents you can enter a monthly count of finished jobs. The app also includes budgets, limits and a stop button for AI traffic you send through CostGrid, with alerts on your Mac.

Download for Mac

View and control

Pricing on request

Everything in View, plus:

  • Pause an agent's Claude keys with one click (OpenAI has no way to pause a key)
  • A monthly cap per agent, checked against the provider's usage report
  • Alerts when an agent is paused or a cap is reached
  • Measured stop time: we time how long a pause takes to work and show you the number
Contact us for a quote

A cap acts on the provider's usage report, which runs behind, so spending stops some time after the cap is crossed, not at once. Tell us about your setup below and we'll send you a quote.

Privacy and key safety

Your data stays on your computer.

Runs on your computer

CostGrid is a Mac app. It works out your costs on your own machine, not on our servers.

Never keeps prompts or AI answers

To notice when the same job is asked again, CostGrid turns the first message of each job into a one-way code on your computer. The words are never saved, the code can't be turned back into them, it never leaves your computer, and it's deleted with your other call records. You can turn this off in Settings.

Keys locked with your Mac's Keychain

Your AI keys are stored encrypted, and the key that unlocks them is kept in the Keychain, where macOS keeps your saved passwords. Your keys are only ever sent to OpenAI or Anthropic, never to us. An admin key you connect goes straight into the Keychain, never into CostGrid's own files, and CostGrid stops using it on the expiry date you chose.

Nothing leaves unless you choose

Nothing is sent from your computer unless you turn it on. With View and control, pausing a key sends Anthropic only that key's id and whether to turn it off or on, and the app shows you exactly that before you press anything.

One key per agent, with an expiry date

We recommend it. You see each agent's cost on its own, you can switch one off without touching the others, and a key with an expiry date stops working on its own.

What this protects against

Your keys aren't sitting readable in plain files, settings or code, and a copied CostGrid folder holds only encrypted keys it can't unlock on another Mac. It can't protect against malware already running on the same Mac; that is out of scope.

FAQ

Straight answers.

Does it read my data?

CostGrid never stores or sends your prompts, your customers' messages or the AI's answers. It saves usage records and, if job matching is on, a one-way code made from each job's first message.

Does it stop hallucinations?

No. CostGrid does not detect hallucinations or check whether an answer is right. It shows what failed and re-run attempts cost you, so you can see which agents waste the most money and fix those first.

Do I need to be technical?

No. Download the app and connect OpenAI or Claude. If a freelancer or someone else set up your AI, CostGrid writes a message you can send them, asking for what it needs.

What does it work with?

OpenAI and Claude today, including agents built in tools like n8n, as long as they run on your own OpenAI or Claude account. To read usage directly, CostGrid needs a key made by someone with an admin or owner role on that account. Without one, drop in the usage file that a developer or billing user can download.

Is there a Windows version?

It's coming. CostGrid is Mac only today.

Is it open source?

No, it's source-available. You can read the code under the Functional Source License (FSL-1.1-ALv2), and each version becomes available under the Apache 2.0 licence two years after its release. The LCSO method behind the numbers is published under CC BY 4.0, so anyone can use it with credit.

Download

Get CostGrid for Mac.

Free to use. No sign-in needed.

Most Macs sold since late 2020 have an Apple chip. Not sure? Click the Apple menu, then About This Mac: it says Apple or Intel next to "Chip" or "Processor".

The first time you open it

CostGrid isn't registered with Apple yet, so macOS will say it can't check the app and stop it from opening. That's expected. To get past it once: open System Settings → Privacy & Security, scroll down to the message about CostGrid, and click Open Anyway. After that it opens like any other app.
Optional

Get a quote or updates by email.

Ask for a View and control quote, or hear first when it and the Windows app are ready. Signing in also opens the full demo. You don't need this to use CostGrid.

Enter a valid email address.

No password: we email you a one-time link. The account is for this website only.