OpenAI Workspace Agents Launch in ChatGPT Powered by Codex

OpenAI workspace agents are now live in ChatGPT, and the pitch is simple: build a bot once, let it run your recurring tasks at 3 a.m. when nobody is at a desk. Announced Wednesday, the new feature runs on OpenAI’s Codex model and is already being tested inside OpenAI’s own teams. It is also a loud signal that the GPTs era, the one where you spent an afternoon tuning a custom chatbot for your sales team, is quietly being folded into something bigger.

What Are OpenAI Workspace Agents?

Workspace agents are persistent assistants that live inside ChatGPT, connect to external apps, remember context across projects, and finish multi-step jobs without being re-prompted every five minutes. That is the pitch. In practice, they sit somewhere between the custom GPTs power users already know and the fully autonomous agents that Silicon Valley keeps promising.

Users spin one up from a new tab in ChatGPT. You describe the workflow you want, ChatGPT maps the steps, wires up the tools, and runs a test. After that, the agent can fire on a schedule or wait for a trigger. Think of it as cron jobs with a language model brain, except Sam Altman wants you to pay by the credit.

The official line from OpenAI workspace agents documentation calls them an evolution of GPTs. That is a polite way of saying GPTs are getting eaten.

Workspace agents are an evolution of GPTs. Powered by Codex, they can take on many of the tasks people already do at work, from preparing reports, to writing code, to responding to messages. They run in the cloud, so they can keep working even when you are not.

Codex Under the Hood and Why That Matters

The choice to plant workspace agents on top of OpenAI Codex is not cosmetic. Codex is the model OpenAI has been sharpening for code generation and structured task execution, which is exactly the kind of plumbing a workspace agent needs when it is stitching together a Slack message, a Google Sheet, and a Jira ticket at 2 a.m.

That is a different job from the polite chat a GPT-4o instance handles in a Dropbox window. Workspace agents have to hold state, call tools, wait for approvals, and pick up the thread hours later. Codex gives them a backbone suited to that rhythm. It also gives OpenAI a clean story to tell developers: if you are already building with Codex, your workspace agent is not a new system, it is the same system with a scheduler attached.

The practical effect is speed. Agents can preflight their own work, test scripts, and hand something usable back to a human reviewer instead of a half-formed draft that still needs three rounds of cleanup.

Who Gets Access and What Does It Cost?

Workspace agents are shipping in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans. Individual Plus subscribers are not on the guest list, which tells you who OpenAI is actually courting here. This is a product built for procurement teams, not hobbyists.

The feature is free until May 6, 2026. After that, OpenAI flips the switch to a credit-based pricing model. The company has not published the full price sheet, which is the part finance departments will be squinting at for the next two weeks. Credit-based pricing usually means you pay for compute, tool calls, and tokens rolled into a single meter, and the meter runs whether a human is watching or not.

For teams already paying Enterprise seat prices, the math is straightforward: if one agent replaces an hour of weekly grunt work across twenty employees, the credits are going to pencil out. If your workflow is vague or your agent hallucinates its way through a quarterly report, the meter still runs.

• ChatGPT Business subscribers
• ChatGPT Enterprise subscribers
• ChatGPT Edu subscribers
• ChatGPT Teachers subscribers

The Prompt Injection Problem Nobody Has Solved

Here is the part the launch post buried. Autonomous agents that read email, scrape the web, and touch internal tools are a soft target for prompt injection attacks, where hostile instructions hide in a document or a webpage and hijack the agent’s next move. OWASP has flagged this as the top risk for large language model applications, and no vendor, OpenAI included, has shipped a clean fix.

OpenAI’s answer is defense in depth. Admins can restrict which data sources and tools an agent can touch. Sensitive actions can be gated behind a human approval step. The platform monitors for prompt injection attempts. That is a reasonable stack of controls. It is not a solution.

The cynical read: OpenAI is shipping the capability and telling enterprises to bring their own risk tolerance. The optimistic read: the only way the industry actually hardens these systems is by deploying them in production under watchful admins. Either way, if you are the security lead at a bank, you are not signing off on workspace agents this quarter.

AI has already helped people work faster on their own, but many of the most important workflows inside an organization depend on shared context, handoffs, and decisions across teams. Workspace agents are designed for that kind of work.

How Does This Fit the Agentic AI Arms Race?

Workspace agents land in the middle of a heavily funded sprint. Google, Microsoft, and Amazon are pouring billions into autonomous systems that can close tickets, triage inboxes, and move money with minimal human babysitting. Microsoft has Copilot agents stitched into the 365 suite. Google is wiring Gemini into Workspace. Amazon has been quieter in the demo videos, louder in the hiring.

OpenAI’s angle is that it owns the model, the chat surface, and now the scheduler. If a workspace agent you build in ChatGPT can also be called from Slack, the operating layer of your day is OpenAI, not Microsoft. That is the land grab nobody is saying out loud.

For the crypto crowd watching this, the overlap is not trivial. Agentic AI frameworks are bleeding into on-chain automation, from trading bots that reason about news flow to DeFi agents that rebalance positions without a human pressing confirm. The rails OpenAI is laying here, persistent context, tool calls, scheduled runs, are the same rails those on-chain agents will eventually piggyback on when the compliance story gets cleaned up.

What Happens to Custom GPTs?

OpenAI insists GPTs are not going anywhere. The company also said it will make it easy to convert a GPT into a workspace agent, which is the kind of line that ends with a product quietly being deprecated twelve months later. Read it how you want.

If you spent the last two years building a library of custom GPTs, the migration path is the story. Converting a chat-shaped GPT into a task-shaped workspace agent is not a one-click exercise. You have to rethink the workflow, decide which steps need human approval, and pick which tools the agent can actually touch. That work pays off only if the agent ends up doing something the GPT could not.

Expect the next wave of ChatGPT product announcements to quietly pull resources away from the GPT store and toward agent marketplaces. The money is in the meter, and the meter is on workspace agents.

This article is for informational purposes only and does not constitute investment advice. Every investment and trading decision involves risk. Readers should conduct their own research before making any financial decisions.