OpenAI's Workspace Agents kill Custom GPTs and take the fight straight to Claude Code

OpenAI launched Workspace Agents on April 22. They’re shared, team-level agents for ChatGPT Business, Enterprise, Edu, plus Teachers that run on Codex, sit behind org-level permissions, keep working in the cloud while you’re asleep, and are invocable from ChatGPT or Slack. They also replace Custom GPTs outright. It’s the clearest signal yet that OpenAI has decided the product category is agent-of-record, not chatbot-with-tools.

What you get today

In OpenAI’s own words, “Teams can now create shared agents that handle complex tasks and long-running workflows, all while operating within the permissions and controls set by their organization.” Translated: your company’s ops lead can build one agent that drafts QBR decks, another that triages bugs, a third that writes security patch release notes. Everyone on the team uses the same agent under the same guardrails. The agent runs when it needs to and reports back in the channel or the ChatGPT thread.

Specifics from the launch:

• Availability. Research preview for ChatGPT Business, Enterprise, Edu, plus Teachers plans as of April 22, 2026.
• Technology. Agents are powered by Codex, the same engine behind OpenAI’s coding agent. That’s the detail that connects this to the bigger Claude Code versus Codex fight.
• Cloud-run, asynchronous. Agents continue executing without your session attached. A long-running report job can finish at 3 a.m. and deposit its output for review.
• Slack integration. Agents are accessible from Slack, so teams don’t have to rehome the work into ChatGPT. For ChatGPT-only shops, the same agent works inside ChatGPT itself.
• Pricing. Free until May 6, 2026. After that, credit-based usage, with details OpenAI is still finalizing.
• Custom GPTs retirement. Workspace Agents replace Custom GPTs, and OpenAI is promising a one-click conversion path for existing GPTs.

How it stacks up against Claude Code and Copilot

OpenAI isn’t shipping this in a vacuum. The agentic-dev and agentic-ops space has been consolidating for a year. Three comparison points are worth calling out.

Claude Code and Anthropic Routines. Claude Code is terminal-native and developer-focused. Routines turn it into a cron runner. The Codex piece OpenAI shipped on Mac desktops in mid-April covers similar ground for interactive coding. Workspace Agents push past coding into general-purpose team work: reports, meeting follow-ups, message drafting. Claude Code has nothing equivalent at the team level yet, though it’s telegraphed multi-seat work.

GitHub Copilot’s recent plan changes. Microsoft and OpenAI moved most of their coding-agent capacity behind the new $39 Copilot Pro+ tier and froze Copilot Pro signups. Workspace Agents give OpenAI a second, independent route to monetize long-running AI work without depending on GitHub. Microsoft will price and meter Copilot. OpenAI will price and meter Workspace Agents. Both share infrastructure but no longer share a single SKU.

Anthropic’s enterprise strategy. Anthropic is betting on Claude Code plus Enterprise plans, and the Figma-rival Claude Design experiment is pushing into verticals. Workspace Agents let OpenAI compete on a different axis: team-wide, cross-function, tied into Slack. A sales ops lead can build a Workspace Agent and share it with her team without ever touching a terminal.

The Codex signal

Buried in OpenAI’s own enterprise rollout post is the line that matters: weekly active Codex users hit 4 million, up from 3 million a few weeks earlier. That’s a 33% jump in a short window, during a period Anthropic has been winning developer mindshare with Opus 4.7 and Claude Code. OpenAI’s read is that Codex is also eating agentic-work outside of pure coding. Workspace Agents are the productized wrapper around that thesis.

Consultancies are piling on. OpenAI named Accenture and PwC as marquee Codex rollout partners, along with Infosys and Cognizant in the second wave. Those are the firms enterprise customers pay to deploy AI systems, and their involvement is the “we’re not losing the enterprise” move. Anthropic already has a parallel Accenture deal on its security product. Both companies are racing to lock up the same systems integrators.

What this means for you

If you run a team already on ChatGPT Business or Enterprise (Edu and Teachers plans are included too), start building agents in the research preview window. It’s free until May 6, which gives you about two weeks to learn the controls, publish a few shared agents, and see which jobs survive into production. The ones that do will feed your team’s collective muscle memory, and by the time credit-based pricing starts you’ll know what’s worth paying for.

If you’re the security or compliance person, the new question is what happens to your Custom GPT inventory. Work with whoever owns your ChatGPT Enterprise tenant to map every Custom GPT to a planned Workspace Agent, figure out which ones need retraining on your data connectors, and set permission boundaries before open enrollment. Permissions on Workspace Agents are organization-scoped, which is a significant step up from the per-user Custom GPT world.

If you’re on a developer team specifically, you now have three overlapping Codex products (Codex CLI/desktop, Workspace Agents, plus Codex Labs) and you’ll have to decide which is the canonical one for different kinds of work. Rough starting split: interactive coding stays in Codex CLI or the Codex desktop app; long-running team workflows go to Workspace Agents; experimental work lives in Codex Labs. That’s not official OpenAI guidance, but it’s how the product surface divides up in practice.

The quietly interesting question is whether Codex weekly-actives continue their current ramp. At 4M and climbing 33% in a few weeks, Codex is on a trajectory that rivals Anthropic’s Claude Code pickup curve. If it sustains, Workspace Agents land into an audience that’s already habituated to Codex. If it stalls (especially after the May 6 credit pricing lands), the product story gets harder even if the feature set stays strong. That’s why the pricing announcement matters more than the feature list: the credit math has to be comparable to Claude Code for team buyers to switch, or comparable enough that the Slack-native experience wins them over by convenience.

One more underappreciated angle. Workspace Agents running in Slack are the closest thing yet to an AI product that doesn’t require you to open a chatbot window. That’s a big deal for a certain kind of buyer. Most enterprise knowledge work already happens in Slack and email. If an agent can DM you the draft of a report that’s been generated while you were at lunch, the cognitive overhead of switching to a new app disappears. OpenAI just made that experience a shipping product.

Why you’re hearing about this now

The timing isn’t random. OpenAI is racing Anthropic for enterprise share during the exact week Google is pitching million-TPU clusters to Anthropic and Microsoft is rebalancing GitHub Copilot pricing. Workspace Agents is the product that fills the gap between “chatbot” and “developer coding tool” for the Fortune 500 buyer who wants a single vendor, a single contract, and agents that live in Slack because that’s where the team already lives.

Watch May 6. That’s when the free preview ends and the credit math starts mattering. OpenAI’s credit-based pricing isn’t priced in public dollars yet, so the first week of May is where this product’s competitive threat to Claude and Gemini for enterprise gets real. If those credits come in noticeably cheaper than Claude Code’s per-seat math, expect Copilot enterprise conversations to stall until Microsoft responds.