ChatGPT Workspace Agents Just Started Billing. Here's What They Actually Cost

OpenAI turned on credit billing for ChatGPT Workspace Agents on July 6, 2026. Until that date, agent runs were free. Now every task an agent completes inside ChatGPT draws down credits on top of your seat price, and the math catches a lot of teams off guard.

Here's the short version. A single agent running 100 tasks a day at default GPT-5 rates costs roughly $300 a month, on top of $20 to $25 per seat. That's before you count the automation you're probably already paying for in Make.com, n8n, or GHL. If you're running a service business and thinking about handing routine work to a ChatGPT agent, you need to know what you're actually buying before you turn it on.

This isn't a "should you use AI" article. You already are. This is about whether a workspace agent belongs in your stack, or whether you're better off with the automation platform you already have.

What Are ChatGPT Workspace Agents, Actually

A workspace agent is a shared, cloud-hosted GPT that runs long tasks on its own and reports back, instead of waiting for you to type each prompt.

Think of it as an evolution of custom GPTs, built on Codex. You give it a job, like drafting weekly reports, triaging a shared inbox, or updating records in a connected tool. It runs in the cloud, so it keeps working after you close the laptop, and it operates inside whatever permissions your workspace admin sets.

That's the pitch. Prepare reports. Write code. Respond to messages. Do the stuff your team already does, minus the person doing it.

For a solo consultant or a 12-person agency, that sounds like leverage. And in some cases it is. But leverage has a meter running on it now, and the meter wasn't public until this week.

Why Is This Happening Right Now

OpenAI isn't the only one moving from single-task assistants to billed, multi-step agents this year, and the shift from demo to daily use is happening across the industry, not just inside ChatGPT.

Enterprise automation platform Alteryx rolled out Agent Studio this month, letting business analysts turn existing data workflows straight into autonomous agents. Insurance tech firm Pace has been running agentic models on live systems for multi-step claims intake, not a demo environment, actual production work. The pattern is consistent. Agents are moving from "answer my question" to "complete this task and report back," and every vendor making that move eventually has to start charging for the compute it takes to complete the task.

That's the context worth knowing. This isn't OpenAI being aggressive with pricing. It's the whole category maturing past the free-trial phase at roughly the same time, which means if you dodge the cost conversation with ChatGPT, you'll likely have it again with the next agent platform you adopt.

What Did the Free Preview Actually Include

Workspace agents ran for free from launch until July 6, 2026, and a lot of teams built workflows during that window without ever seeing a bill.

That's the trap. If you spun up an agent in May or June to auto-draft client updates or pull data into a report, you built a habit around a $0 tool. As of Monday this week, that habit has a price tag, and it's billed in credits, not dollars, which makes it harder to eyeball.

Runs triggered from inside ChatGPT now draw down workspace credits. Runs triggered from outside ChatGPT, like an agent replying inside a Slack channel, are still in free preview for now. That distinction matters more than it sounds like it should, because it means where you trigger the agent changes what you pay.

How Much Does an Agent Run Actually Cost

A typical agent run costs 5 to 25 credits, and credits are priced against token usage, not a flat per-task fee.

Here's a real example from OpenAI's own rate card. A GPT-5.5 workspace agent run using 20,000 input tokens, 80,000 cached input tokens, and 5,000 output tokens costs about 7.25 credits. Input runs 125 credits per million tokens. Cached input runs 12.50 credits per million tokens. Output runs 750 credits per million tokens, because generation is the expensive part.

Output is where the cost hides. An agent that reads a lot and writes a little is cheap. An agent that drafts long reports, writes code, or generates detailed client deliverables burns through credits fast, because every word it produces costs roughly 60 times more than every word it reads.

Run the numbers on a moderately active agent, one handling 100 invocations a day, and you land around $300 a month at default GPT-5 rates. That's one agent. If you're picturing three or four agents across sales, ops, and delivery, you're looking at four figures a month before you've automated a single new thing you couldn't already do with a $20 tool.

What Does a Business Actually Pay to Get Access

ChatGPT Business runs $20 per seat monthly on annual billing, or $25 monthly, with a two-seat minimum, and workspace agent credits sit on top of that.

Enterprise starts around $60 per seat with a 150-seat minimum, and adds SCIM, EKM, domain verification, and role-based access controls. For most service businesses under 20 people, Business is the relevant tier, not Enterprise.

So the real cost isn't the seat. It's the seat plus whatever your agents burn through in credits, and OpenAI gave teams roughly 26 days of notice between announcing credit pricing and turning it on. If nobody on your team was tracking that, you found out this week the hard way, in your bill.

Is a ChatGPT Agent Cheaper Than the Automation You Already Have

For most repeatable, rules-based workflows, Make.com, n8n, or Zapier will beat a ChatGPT workspace agent on cost by a wide margin.

Compare the two models directly. Make's Core plan runs $9 a month for 10,000 operations. Standard is $29 for 40,000 operations. A 10-step Make scenario costs 10 operations per run, so even the Core plan covers 1,000 full workflow runs a month for nine dollars.

n8n's cloud Starter plan is €20 a month for 2,500 executions, where an entire multi-step workflow counts as a single execution, not a per-step charge. Self-hosted n8n runs about £20 a month in server costs and is the cheapest option at real scale, because you're paying for compute, not per-task billing.

Zapier is the expensive end of that comparison. Its Starter plan is £20 for 750 tasks, and each step in a workflow counts as a separate task. Run a 10-step workflow 10,000 times a month on Zapier and you're near £940. Run the same workload on n8n and you can cut that by 80 to 90 percent, because n8n bills per execution, not per action.

Now put a ChatGPT agent next to that. If your workflow is genuinely rules-based, like "when a form is submitted, create a CRM record, send a Slack alert, and add a calendar event," that's a job for Make or n8n, full stop. You're not paying for reasoning. You're paying for reliable, cheap execution, and a $9 to $50 monthly plan already does it.

When Does a ChatGPT Agent Actually Make Sense

A workspace agent earns its cost when the task requires judgment, not just steps, like drafting a first version of a client proposal or summarizing a messy inbox before a human reviews it.

This is the test I use with clients, and I call it the Rent vs. Build Rule. It has three questions:

First, does the task require language understanding or judgment, not just moving data from A to B? If yes, that's an agent's job. If no, that's an automation platform's job.

Second, is the output reviewed by a human before it goes anywhere important? If the agent is drafting, not deciding, the cost of an occasional bad output is low, and the agent earns its credits.

Third, would a fixed monthly automation plan already cover this at a fraction of the cost? If Make or n8n can do it with a webhook and a filter, don't pay per-token pricing for a job that doesn't need language reasoning.

A workspace agent triaging a shared support inbox and drafting first-pass replies for a human to approve is a good use of the Rent side. An agent that just moves a lead from your form into your CRM is not. That's a $9 Make scenario wearing an AI costume.

What Does This Look Like for an Actual Service Business

Run the math on a 6-person marketing agency and the gap between "renting" and "building" gets easy to see in real dollars.

Say that agency puts a workspace agent on client reporting. It drafts a performance summary for each client once a week, reads campaign data, and writes a paragraph of commentary. That's language-heavy work, real output, so it sits on the Rent side. At 20 clients, one report a week, that agent might run 80 to 100 invocations a month. At $300 for 100 invocations a day, a monthly cadence like that lands closer to $30 to $60 a month in credits, plus the seat. Reasonable. The agent is doing something a $20 automation tool can't.

Now say that same agency also has an agent that watches a shared inbox and forwards new lead emails into their CRM with a tag. No drafting, no judgment, just read and route. Priced the same way, that's paying premium-token rates for a job a $9 Make scenario handles instantly, with a webhook and a router. Two agents, same platform, wildly different math, because only one of them needed language reasoning to do its job.

This is the piece most teams miss when they adopt a new tool. It's not "ChatGPT agents good" or "ChatGPT agents bad." It's that pricing by the token means every task needs to earn its place, and a lot of tasks people are routing through an LLM never needed one.

What Should You Actually Do This Week

Audit which agent runs you built during the free preview, tag each one as judgment-based or rules-based, and move the rules-based ones back to your existing automation platform before the credits add up.

Most teams that built agents in May and June never separated the two. They used ChatGPT because it was free and easy to prototype in, not because it was the right long-term tool. That's fine. Prototyping in the free tool is smart. Leaving it there after the meter starts running is not.

Pull your list of active workspace agents. For each one, ask the three Rent vs. Build questions. Anything that's pure data movement, no judgment required, goes to Make, n8n, or your GHL workflows, where it'll run for a fraction of the price. Anything that requires actual reading, drafting, or reasoning stays on the agent, and you budget for it like the real line item it now is.

The teams that get burned by this aren't the ones using AI agents. They're the ones who never checked which agents were doing judgment work versus busywork, and got a July bill that made the difference very clear.

If you want a clear picture of what AI can actually do for your specific operation, book a free AI Clarity Call. Thirty minutes, no pitch, you leave with a real answer.

If you want to learn alongside other operators and stay current on what is working, join the Abra AI community. That is where I share what I am actually building.

Subscribe to the newsletter for more breakdowns like this.