On April 22, 2026, OpenAI announced ChatGPT workspace agents — a new feature that replaces custom GPTs with shared, Codex-powered AI agents that take actions, run in the cloud, and live inside Slack alongside ChatGPT. The openai workspace agents rollout is in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans, free until May 6, 2026, after which a credit-based pricing model takes over.
This is the biggest product change OpenAI has made to ChatGPT for teams since custom GPTs launched in late 2023. Here is a complete breakdown of what shipped, what it costs, what it does, and how it compares to the alternatives.
What Are ChatGPT Workspace Agents?
ChatGPT workspace agents are shared AI agents that organizations can build once and use across a team. They run in the cloud, take multi-step actions across connected tools, retain memory across sessions, and can be triggered on schedules or deployed inside Slack. Positioned as ChatGPT agents for teams rather than individual productivity tools, they’re built around the idea that most meaningful work happens across handoffs, not inside a single person’s chat window.
OpenAI positions them as the successor to custom GPTs. Where a custom GPT was essentially a saved system prompt with a better interface, a workspace agent is a persistent piece of software that can execute real workflows — drafting reports, responding to messages, pulling data, filing tickets, or running month-end processes without constant prompting.
The underlying engine is Codex, OpenAI’s coding model, now doing double duty as the reasoning layer for persistent agents. These codex workspace agents can actually execute workflows instead of just describing them — the shift from “tool that answers questions” to “tool that completes tasks.”
Three behaviors separate workspace agents from custom GPTs:
- They take actions across connected tools — drafting emails, updating documents, filing tickets, posting in Slack channels.
- They run without you — scheduled jobs, event triggers, and Slack-based interactions mean an agent can keep working when the user closes ChatGPT.
- They have memory and improve over time — corrections persist, so an agent your team trained in March produces better output in April.
Who Can Access Workspace Agents? ChatGPT Business Agents and Eligible Plans
Workspace agents are available in research preview for the following ChatGPT plans:
- ChatGPT Business (starts at $20 per user per month)
- ChatGPT Enterprise (custom pricing)
- ChatGPT Edu
- ChatGPT Teachers
Workspace agents are not available on ChatGPT Plus, ChatGPT Pro, or the free tier. OpenAI has not published a timeline for extending access to individual plans.
If you already have an eligible plan, the Agents tab is live in the ChatGPT sidebar. Click it, describe a workflow your team does often, and ChatGPT walks you through turning it into an agent.
ChatGPT Workspace Agents Pricing: What It Costs After May 6
This is the most important line in the launch materials, and most coverage has skipped it. Here’s the workspace agents pricing breakdown:
- Free until May 6, 2026 — research preview period, no additional charge beyond your existing ChatGPT Business or Enterprise subscription.
- After May 6 — OpenAI moves to a credit-based pricing model. Exact rates have not been published.
Credit-based pricing usually means you pay per action, per tool call, or per minute of compute time. Teams that deploy several scheduled agents across multiple connected tools should budget for this to cost meaningfully more than the base $20 per user per month once real workloads scale.
If you are evaluating whether to adopt workspace agents, the practical move is to build and test during the free window, then reassess in early May once the pricing structure is confirmed.
Which Tools Do Workspace Agents Connect To? Slack, Salesforce, and More
Out of the box, ChatGPT workspace agents connect to the standard enterprise stack:
- Slack (both as a surface for deploying agents and as a data source)
- Google Drive
- Microsoft 365 (Word, Excel, Outlook, Teams)
- Salesforce
- Notion
- Atlassian Rovo (Jira, Confluence)
OpenAI has indicated that additional connectors are coming. The connector library is the difference between an agent that sounds smart in a demo and one that’s actually useful on Monday morning — a team building a weekly metrics agent needs access to the tools where the metrics actually live.
Slack deployment is the connector worth paying closest attention to. The workspace agents Slack integration lets teams add agents directly to channels, respond to messages, and participate in conversations where work happens — without requiring anyone to open ChatGPT.
What Can You Actually Build?
OpenAI’s launch materials highlight five reference workflows. These aren’t theoretical — they describe the shape of agent that’s most likely to deliver real value today:
1. Weekly Metrics Reporting Agent
Pulls data every Friday from connected tools, generates charts, drafts narrative summaries, and delivers the report to Slack or a shared document. Replaces the 2–4 hour manual process most marketing and ops teams still run every week.
2. Product Feedback Routing Agent
Monitors Slack channels, support tools, and public review sites, clusters feedback into themes, and surfaces the top priorities to product and marketing teams. Turns scattered signals into a weekly action list.
3. Software Review Agent
Triages employee software requests, checks against internal policy, routes approvals, and files IT tickets. Internal operations workflow that eats manager hours in mid-sized companies.
4. Lead Outreach Agent
Researches inbound leads using firmographic data, scores them, drafts follow-up emails, and updates CRM records. Compresses pre-call prep from 15 minutes to 90 seconds.
5. Month-End Close Agent (Finance)
Prepares journal entries, reconciles balance sheets, runs variance analysis, and generates workpapers — following internal policies. OpenAI’s own accounting team uses one of these internally.
For marketing teams specifically, the weekly metrics reporter and the lead research agent are the fastest routes to measurable time savings. If your team has already explored AI marketing automation tools, workspace agents consolidate several of those use cases into a single persistent agent.
Governance, Admin Controls, and Security
One of the key reasons workspace agents are positioned as an enterprise product is the governance layer built into the system. Admins can:
- Control who in the organization is allowed to build agents
- Control which tools and data sources an agent can access
- Require human approval for sensitive actions (sending emails, modifying records)
- Monitor usage and performance across agents
- Manage sharing permissions at the agent level
This matters more than it sounds. Agents that can take actions across your tools are, by definition, a security surface. OpenAI has been explicit about the prompt injection risk — an attacker embedding malicious instructions inside a document, email, or web page that an agent processes, potentially hijacking the agent’s behavior.
The mitigations in the product are the right ones: scope what tools an agent can reach, require approval for sensitive actions, and monitor patterns. But teams deploying workspace agents still need to treat them like any other piece of production software. No agent should have write access to customer data or financial systems without a human approval step in the workflow.
How Do Workspace Agents Compare to Alternatives?
ChatGPT workspace agents are not the only product in this category. Four other major launches over the past eight months have aimed at the same problem:
- Microsoft Copilot Studio — deeply integrated with Microsoft 365, strongest if your team lives in Outlook, Teams, and SharePoint
- Google Agentspace — Workspace-native, strongest on Gmail, Drive, and Google-first stacks
- Salesforce Agentforce — Salesforce rebuilt itself around agents; strongest for revenue teams already on the Salesforce platform
- Anthropic’s Claude Managed Agents — Claude-powered, strong on reasoning-heavy tasks, paired with the Claude for Work platform
The choice between these is less about model quality and more about where your team already works. A SaaS startup running on Slack, Notion, and HubSpot has different gravity than an enterprise on Microsoft 365 and Salesforce.
For most B2B SaaS teams under 50 people, ChatGPT workspace agents and Claude Managed Agents are the two realistic options. Workspace agents currently have the edge on Slack-native deployment and the breadth of out-of-the-box connectors.
What Happens to Custom GPTs?
OpenAI’s official position: custom GPTs remain available during the workspace agents research preview. A conversion path from GPTs to workspace agents is planned.
Practical reading of the signals: all new product investment is going into agents. Every example in the launch post frames what agents can do that GPTs can’t. Enterprise customers are being guided toward migration.
If your team currently uses custom GPTs, nothing breaks immediately. But for any new build — every “we should make a GPT for this” conversation — the answer is now workspace agents.
How to Get Started This Week
For teams on an eligible ChatGPT plan:
- Open the Agents tab in the ChatGPT sidebar
- Describe a repeated team workflow in natural language
- Let ChatGPT walk you through tool connections, approvals, and testing
- Start with one agent, not five — the weekly metrics reporter is the easiest to QA because you already know what good output looks like
- Run it for three weeks before building a second agent
- Document every correction you make — that’s the training data that improves the agent over time
- Set a May 5 calendar reminder to review pricing before the free period ends
For teams not on an eligible plan, the recommendation is to wait. Pricing will be clearer after May 6. Individual plan support may come. The general-purpose AI marketing tools in your stack still cover most of the ground workspace agents would cover for a small team.
The Bigger Picture
The 2023–2025 AI era looked like this: write a prompt, get a response, copy it somewhere. The 2026 era looks like this: software that does work, not software that gives you answers. Workspace agents, Claude Managed Agents, Copilot Studio, Agentspace, Agentforce — they’re all pointing at the same destination from different starting points.
For B2B SaaS teams, the honest read is that agents will be unreliable on judgment-heavy tasks for at least another 6–12 months. They’ll be genuinely useful for tasks that involve compilation, routing, drafting, and scheduled reporting — which is where most marketing ops work actually lives anyway.
If your team spends Fridays pulling the same numbers from the same tools into the same deck, that’s the work workspace agents are built to take off your plate. That’s the productivity gain. Not “AI runs your marketing.” More like: “your people stop doing the boring middle.