Custom GPTs Are Dead. Here's What Replaces Them.

On April 22, OpenAI published Introducing workspace agents in ChatGPT and pushed every business that ever built a Custom GPT into a two-week decision window. Workspace Agents are the official successor. They’re free until May 6. After that, credit-based pricing starts, and the clock on your existing GPT library quietly starts ticking.

If your team built even one Custom GPT that a human actually uses every week, this affects you. The migration path is announced but not yet easy. The pricing is announced but not yet specific. And the capability gap between what a Custom GPT does and what a Workspace Agent does is wide enough that “just convert them” is the wrong frame.

Here’s the decision framework I’d run this week, before the free window closes.

Quick Verdict

Decision Point	What Changed on April 22
Custom GPTs status	Still usable today; OpenAI has said business/enterprise/edu/teachers users will be required to migrate at a yet-to-be-determined future date (VentureBeat)
Replacement product	Workspace Agents, powered by Codex
Native integrations	Slack, Google Drive, Microsoft apps, Salesforce, Notion, Atlassian Rovo (VentureBeat)
What agents do that GPTs don’t	Run on schedules, hold memory across sessions, improve through correction, take actions in third-party apps
Free trial window	Free for Business, Enterprise, Edu, and Teachers plans until May 6, 2026
Pricing after May 6	Credit-based. Structure not yet disclosed.
Base plan cost	ChatGPT Business starts at $20 per user per month
What to do this week	Inventory GPTs, pick one to rebuild as an agent, measure the delta before May 6

What a Workspace Agent Actually Is

A Custom GPT was a prompt wrapper. You gave it instructions, uploaded a few files, maybe attached a couple of actions, and it responded when someone opened the chat. It was useful. It was also passive. The moment the user closed the tab, the GPT stopped existing.

A Workspace Agent is the opposite. OpenAI describes them as “an evolution of GPTs,” which is marketing copy for three actual changes:

1. They run in the cloud on their own schedule. The agent keeps working when you’re not at your desk. A weekly pipeline review agent runs Monday morning whether you log in or not.
2. They connect directly to your work systems. Native integrations with Slack, Google Drive, Microsoft apps, Salesforce, Notion, and Atlassian Rovo mean the agent isn’t asking you to paste data in. It’s reading from the source and writing back.
3. They hold memory and accept correction. An agent that drafted your weekly report wrong can be corrected conversationally, and the correction sticks. The next week’s draft reflects the change. Custom GPTs didn’t do this.

The engine is Codex, which is how they execute multi-step work rather than just replying. According to 9to5Mac, agents handle “preparing reports, writing code, and responding to messages” as long-running workflows that span hours or days.

That is a different product category than a Custom GPT. Treating it as a rename is how you miss the migration.

Why the Timing Is Not Coincidental

OpenAI didn’t announce a replacement for Custom GPTs in a vacuum. Every major AI vendor shipped an agent platform between January and April 2026. Microsoft pushed agents deeper into Teams and Office. Salesforce kept iterating Agentforce. Anthropic launched the Claude partner network and its Agent Skills self-serve layer. Google moved Gemini toward workflow automation.

The GPT Store, which peaked as a concept in early 2024, was already a side quest by mid-2025. Nobody was building a durable business on it. OpenAI knows this. Workspace Agents is the grown-up version of that bet, aimed at the part of the market that actually pays: enterprises with real workflows across Slack, Salesforce, and Notion who need agents that do work inside their stack.

Everything I wrote in my piece on what OpenAI fixing the biggest agent blocker means applies here. The blocker was never capability. It was integration and persistence. Workspace Agents ship both in the same release.

Custom GPT vs Workspace Agent: The Real Comparison

Here’s the side-by-side most migration posts are skipping.

Capability	Custom GPT	Workspace Agent
Runs on a schedule	No	Yes
Memory across sessions	Limited	Yes, with correction
Native Slack integration	No	Yes
Salesforce integration	Via custom actions	Native
Google Drive, Microsoft apps	Read-only in most cases	Read and write
Notion, Atlassian Rovo	No	Native
Shared across a team	Yes	Yes
Improves through user correction	No	Yes
Works in background while you’re offline	No	Yes
Takes actions in third-party apps	Limited	Core function
Pricing	Included in plan	Credit-based after May 6

The columns that say “No” on the left and “Yes” on the right are the ones that matter. A Custom GPT that drafted your weekly client update required a human to open it, paste context, copy the output, and send. A Workspace Agent reads the client’s Slack channel and Salesforce record on Friday morning, drafts the update in the agent’s thread, and pings the account owner for a one-click approval.

That is a different product, not a newer version of the same product. Your migration plan should start from that assumption.

What Will a Workspace Agent Cost You?

OpenAI was explicit about the trial and vague about the price. Here’s what we know:

• Free in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans until May 6, 2026
• Credit-based pricing starts May 6
• ChatGPT Business base plan is $20 per user per month, per the ChatGPT Business pricing page
• The credit rate per agent run has not been disclosed

“Credit-based” is the part CFOs should pay attention to. Per-user licensing is predictable. Credit-based pricing is not, and it’s the model most likely to produce a surprise bill when an agent starts triggering itself on a schedule you forgot about.

This is the same structural risk I wrote about in the agent sprawl prevention piece. Per-run pricing means every scheduled trigger is a billable event. A team that rebuilds six GPTs as agents and sets each one to run daily is looking at 30 runs a week per agent, 180 runs a week across the team, before a single human prompts anything. Multiply by the credit cost OpenAI hasn’t announced yet, and you have a line item nobody budgeted.

The two-week free window is the time to estimate that number. After May 6, you’re estimating with real invoices.

The Two-Week Migration Framework

If your team has shipped Custom GPTs that people actually use, here is the framework I’d run between now and May 6. It fits in a week if someone owns it.

Step 1: Inventory Every Custom GPT In Use

Open your workspace’s GPT list. Pull usage data from the last 90 days. For each GPT, write down three things:

• Weekly active users
• The workflow it supports (not what it does — what job it’s part of)
• Whether the workflow touches Slack, Google Drive, Salesforce, Microsoft apps, Notion, or Atlassian Rovo

Most teams find out here that half the Custom GPTs they built are used by two people a month. Kill those first. They’re not worth migrating, and they were never worth the licensing noise.

Step 2: Rank by Agent Upside

For the GPTs that survived step 1, score each one on a simple three-point scale:

High upside. The workflow runs on a predictable schedule, touches a system in the native integration list, and a human currently wastes time as a courier between the tool and the chat. Example: a sales prep GPT that needs a rep to paste in Salesforce data every morning.

Medium upside. The workflow is ad hoc but frequent, and a memory-holding agent with correction would measurably reduce the prompt load. Example: a brand voice editor that three writers use daily.

Low upside. The workflow is genuinely one-shot, rare, or fine as a pure prompt. Example: a resume-review GPT used once a quarter.

Only the “high upside” column is worth rebuilding in the free window. Medium is a next-quarter problem. Low stays as a GPT until OpenAI forces conversion.

Step 3: Rebuild One High-Upside GPT as an Agent

Pick the single highest-upside workflow and rebuild it as a Workspace Agent this week. Not six. One. The point is to learn the real migration cost and the real value delta before you commit to the other five.

What to measure during the rebuild:

• Time to set up the integrations (Slack, Salesforce, Drive) versus the old GPT
• Time from prompt to result on a scheduled run versus the manual GPT cycle
• Error rate on outputs the first day, week, and two weeks in
• How often the agent needs correction, and whether the correction sticks

Keep the original Custom GPT running in parallel. You will need the comparison data when you defend the budget line after May 6.

Step 4: Estimate Your Credit Burn Before May 6

Once your one agent has run for a full week, extrapolate. Count every scheduled trigger, every user-prompted run, every action taken against Slack or Salesforce. That’s your per-agent weekly volume. Multiply by the number of high-upside agents you plan to ship.

Compare that to your current ChatGPT Business per-seat spend. The gap is what your credit-based bill could look like at the pricing OpenAI hasn’t announced yet. Present the range to finance before you commit. If the number makes the CFO wince, you don’t have a migration problem. You have an agent-count problem, and the fix is discipline, not licenses.

Step 5: Decide What to Kill, What to Keep, What to Rebuild

By the end of the free window, you should have four buckets:

1. GPTs to retire. Low-usage, low-upside. Turn them off.
2. GPTs to keep as GPTs. Workflows that are genuinely chat-shaped and don’t need scheduling, memory, or integrations. They stay until OpenAI forces conversion.
3. GPTs to rebuild as Workspace Agents now. The high-upside set. Migrate in the second week of May.
4. GPTs waiting on clarity. Anything where the credit pricing changes the ROI math. Park these until OpenAI publishes the actual per-run rate.

The teams I’d expect to do this well are the same ones who already ran a consolidation of their AI tool stack. The teams still adding tools without a budget are going to migrate everything, set every agent on a daily schedule, and discover their May invoice in June.

What Workspace Agents Still Can’t Do

No vendor announcement survives contact with a real workflow. Here is where Workspace Agents, as of April 23, will bite you:

Credit pricing is a black box. OpenAI has not published a per-run rate. Any ROI calc you do before that number lands is a guess. Run the pilot, but don’t commit the rest of the migration until the price is visible.

Integrations are native, not deep. “Native Slack integration” means the agent can read and post. It does not necessarily mean the agent can handle complex Slack workflow rules, retention policies, or enterprise security controls at the detail level. Ask your security team to audit before you point an agent at a sensitive channel.

“Improves through correction” is not self-learning. Don’t confuse this with autonomous improvement. The agent updates behavior when you correct it. It does not reason about failures it wasn’t told about. Your error rate drops because a human is still in the loop, not because the model woke up smarter.

Migration tooling isn’t here yet. OpenAI has promised a conversion path from GPTs to agents. That path does not exist today. Rebuilding a GPT as an agent this week is a manual job. Plan accordingly.

I wrote about the general version of this pattern in stop building AI agents. Every vendor-driven agent shift requires the same discipline: pilot one, measure, decide, then scale. Skipping the pilot because the vendor announcement felt urgent is how you get the bill without the value.

The Part That Nobody Is Writing Down

The hidden implication of this launch is that the GPT Store was always a placeholder. OpenAI’s real enterprise bet is not “everyone builds a custom chatbot.” It’s “every team runs a portfolio of agents that operate across their existing stack, priced by consumption.”

That business model favors vendors. It favors OpenAI. It does not automatically favor buyers.

The counterweight is discipline on the buyer side. Pick the three workflows per team that are genuinely improved by an agent versus a prompt. Put a named owner on the agent. Measure weekly. Kill what doesn’t earn its credits. This is the same three-slot ceiling I argued for in the AI productivity trap piece. Tools and agents obey the same rule. More is not better. The right three are.

The 20% of firms capturing the real value from AI are going to run this migration like a budget decision. The other 80% are going to convert everything, set every agent to daily, and have a painful May.

Your Move This Week

Do three things before the free window closes.

1. Inventory your Custom GPTs. Pull usage data. Kill the dead ones. Rank the rest.
2. Rebuild your single highest-upside GPT as a Workspace Agent. Measure the time, error, and correction delta against the old GPT.
3. Model your credit burn with placeholder pricing. Present the range to finance before May 6 so the conversation is about tradeoffs, not surprises.

Custom GPTs are not technically dead — individual users can keep theirs for the foreseeable future — but OpenAI has confirmed business and enterprise GPTs will be required to migrate on a date they haven’t named yet. What’s dead is the idea that a GPT is the endpoint of a workflow. The endpoint, starting now, is an agent that runs across your stack, holds memory, takes actions, and charges you per run for the privilege.

Pick the ones worth the credits. Retire the rest. The two weeks you have are enough to do that well. After May 6, the same migration costs real money.

---

Related Reading:

• OpenAI Fixed the Biggest Agent Blocker. Now What?
• Stop Building AI Agents: The 2026 Licensing Reality
• Agent Sprawl: An SMB Prevention Guide for 2026
• Your AI Tools Are Creating Work, Not Saving It
• Your AI Tool Stack Has an Expiration Date