ChatGPT workspace agents graduate from free preview to credit-based pricing today, May 6, 2026, marking the moment OpenAI‘s shared-team agent infrastructure becomes a real budget line for enterprise buyers. Workspace agents let teams create durable, multi-step AI workers that operate inside an organization’s ChatGPT environment — handling research, document drafting, data analysis, and complex workflows that span tools and time. Unlike individual GPTs or Custom Instructions, workspace agents are first-class organizational citizens with permissions, audit trails, shared ownership, and the operational primitives enterprises actually need. This is the practical breakdown for the IT leaders, ops teams, and developers evaluating workspace agents for production deployment.
What’s actually new
Workspace agents introduce three things that ChatGPT Enterprise didn’t have before. First, shared agency: a single agent definition that any authorized team member can invoke, kick off, monitor, and edit. Earlier ChatGPT customizations (Custom GPTs) were per-user; workspace agents are per-team. The shared layer matches how enterprise work actually happens — a finance ops agent isn’t owned by one analyst, it’s owned by the finance team.
Second, long-running execution with checkpointing. A workspace agent can run for hours or days. It pauses for human input when needed (approval gates, ambiguous decisions, missing data), persists state across pauses, and resumes from checkpoints when execution continues. This is the durability layer that makes agents useful for workflows that don’t complete in a single response — quarterly reporting, multi-step research projects, batch document generation, complex customer outreach.
Third, permissions and observability. Workspace agents inherit the organization’s existing identity and access controls — SSO, group membership, data-source permissions. Every agent invocation produces an audit log. Spend per agent is tracked and budgetable. The compliance and operational primitives that ChatGPT Enterprise customers expect for human users now extend cleanly to AI agents.
The pricing change matters. Until today, workspace agents were free during a beta period. Starting May 6, 2026, agents draw from a credit-based pricing pool: each invocation consumes credits proportional to compute (tokens), tools used (web search, file reads, code execution), and any external API calls the agent makes. Workspace administrators set per-agent budgets, per-user limits, and total monthly caps. The cost model is similar to traditional cloud-resource billing — predictable when usage is bounded, variable when it scales.
The capability profile of workspace agents draws from OpenAI’s existing ChatGPT agent product (announced earlier in 2026), but with shared-team and enterprise-grade twists. Each agent can use the same toolset: web search, file reads from the workspace’s connected data sources, code execution in a sandbox, image and document generation, and integrations with hundreds of third-party services through the existing ChatGPT connectors framework. Agents inherit the workspace’s data access, so an analytics agent can read the analytics-team’s connected data sources without having to be granted separate permissions.
Notably, workspace agents are also the first ChatGPT product to formally support cross-workspace coordination. An agent in your workspace can call an agent in a partner workspace (with explicit authorization on both sides). This is OpenAI’s first concrete step toward inter-organizational agent ecosystems — a pattern that several startups have been working toward via separate protocols.
Why it matters
- Enterprise AI shifts from “individual assistant” to “shared infrastructure.” The next 18 months of enterprise AI adoption will be defined by shared agents that coordinate work across teams, not by individual employees using personal AI assistants. Workspace agents are the platform play that enables this transition inside ChatGPT.
- Pricing transparency forces ROI conversations. Free preview let teams build agents without scrutinizing cost. Credit-based pricing forces explicit ROI calculations: this agent saves $X/week and costs $Y in credits — is the ratio acceptable? Most teams will discover surprising patterns about which agents pay back and which don’t.
- The audit-trail story closes a long-standing enterprise gap. ChatGPT Enterprise had logs but no way to attribute actions cleanly to a specific shared agent. Workspace agents fix this. Compliance and security teams that were skeptical of enterprise ChatGPT adoption now have the audit primitives they needed.
- OpenAI competes with LangGraph and similar frameworks at the platform layer. Workspace agents do for ChatGPT customers what LangGraph does for self-hosted deployments — durable, stateful, observable agentic workflows. Customers who previously had to choose between ChatGPT (high quality, weak orchestration) and LangGraph (good orchestration, BYO model) can now get both inside ChatGPT.
- The credit-based pricing model presents a budget challenge. Unlike per-user seat licensing, credit-based pricing scales with usage. Teams that build heavily-used agents face variable monthly bills. Budget governance becomes more important than ever.
- Cross-workspace coordination opens a new platform vector. If workspace agents in one organization can call workspace agents in another, OpenAI is positioning ChatGPT to be the substrate for inter-organizational AI workflows. This is a long-term strategic play that may not materialize quickly but is worth tracking.
How to use it today
Workspace agents are available now in ChatGPT Enterprise, Business, and Education tiers. Pro tier gets a limited preview. The setup workflow takes 30-60 minutes for a first agent.
- Confirm tier eligibility. Workspace administrators check their plan in Settings → Billing. Workspace agents require Business, Enterprise, or Education. Upgrade if needed.
- Create the agent. Navigate to Workspace → Agents → Create New. Name the agent, write the system prompt that defines its role and behavior, choose the tools it can access (web search, code execution, connected data sources), and set its permission scope (which users in the workspace can invoke it).
MASK12
- Configure permissions and budget. Set who can invoke the agent (specific users, groups, or all workspace members), what data sources it can access, and the monthly credit budget. Default permissions are restrictive; expand intentionally.
- Test in a controlled scope. Invoke the agent yourself with a representative task. Review the trace — what tools did it use, what was the cost, what was the output quality? Iterate on the system prompt until results meet your bar.
- Roll out to your team. Document how to invoke the agent (via @-mention in conversations, dedicated agent panel, or API). Train heavy users on best practices: clear instructions, structured output expectations, when to escalate to a human.
- Programmatic invocation via the API. For workflows that should run automatically (scheduled reports, event-triggered tasks), invoke the agent via the OpenAI API:
MASK13
- Monitor usage and cost. The Workspace Admin dashboard shows agent invocations, credit consumption per agent, error rates, and average task duration. Establish a weekly review cadence; agents that drift from expected cost should be investigated.
- Iterate based on real usage. System prompts that worked in testing rarely survive contact with diverse real-world tasks. Plan for ongoing prompt refinement, especially in the first 30 days. Track which kinds of tasks the agent handles well versus poorly; expand or constrain its scope accordingly.
How it compares
| Platform | Shared agents | Long-running tasks | Audit and permissions | Pricing model | Best for |
|---|---|---|---|---|---|
| ChatGPT Workspace Agents | Yes (team-scoped) | Yes (durable, hours-to-days) | Strong, inherits Enterprise SSO | Credit-based | ChatGPT Enterprise customers |
| Anthropic Orbit (in Claude Cowork) | Yes (team-scoped) | Continuous + on-demand | Strong, typed connectors | Per-seat ($30 base + $10 Orbit) | Claude Cowork users |
| Microsoft Copilot Studio | Yes (tenant-scoped) | Limited; works best for shorter tasks | Strong, inherits Entra ID | Per-seat + Copilot Studio licensing | Microsoft 365 customers |
| Google Gemini Workspace | Limited (mostly per-user) | Reactive primarily | Strong, inherits Workspace IAM | Per-seat $30/mo | Workspace customers |
| LangGraph (self-hosted) | Yes (BYO multi-tenant) | Yes (durable execution) | Strong (BYO architecture) | Open-source + LLM API costs | Teams wanting full control |
The competitive picture: workspace agents put OpenAI in direct competition with Anthropic Orbit on the shared-team-agent dimension and with Microsoft Copilot Studio on the platform-integration dimension. For ChatGPT Enterprise customers, workspace agents are now the obvious next step — same data, same identity, same admin tooling. For mixed environments, the choice depends on which AI provider you’re already deepest with.
One implicit takeaway: pricing models differ meaningfully. Per-seat pricing (Orbit, Copilot, Gemini Workspace) is predictable but doesn’t reflect actual usage. Credit-based pricing (ChatGPT workspace agents) reflects usage but creates budget volatility. Self-hosted (LangGraph) bills only the underlying LLM API plus your operational cost. The right model depends on your organization’s preferences for predictability versus efficiency.
What’s next
Three trajectories worth watching for the rest of 2026.
Cross-workspace coordination matures. The current cross-workspace primitive is rudimentary; expect richer protocols as customer demand surfaces. The interesting use cases — supplier coordination, vendor management, cross-organizational research — require trust frameworks, dispute resolution, and identity standards that don’t yet exist. OpenAI is positioning early; the technology will catch up over 12-24 months.
Pricing optimization becomes a discipline. Credit-based pricing forces customers to think about agent efficiency. Expect a wave of best-practice guidance, third-party tooling for agent cost analysis, and OpenAI-native features for budget governance. The “FinOps for AI agents” category emerges as a real specialty.
Agent quality and safety controls deepen. Workspace agents at enterprise scale produce edge cases that the OpenAI team will encode into platform features: better evaluation tooling, output validation, behavioral guardrails, and incident-response support. The platform’s safety story improves with usage; expect significant updates through 2026 H2.
The deeper observation: enterprise AI in 2026 is converging on a small number of platform plays. ChatGPT, Microsoft Copilot, Anthropic Cowork, and Google Workspace are the four credible options for most enterprises. Within each, shared-team agents are becoming the defining capability. Customers will pick a primary platform and integrate it deeply rather than running multiple platforms in parallel. ChatGPT Workspace Agents is OpenAI’s strongest play yet to be that primary platform for the enterprise customer base they’ve already won.
Frequently Asked Questions
What’s the difference between workspace agents and Custom GPTs?
Custom GPTs are per-user customizations of ChatGPT — your personal helper that follows your custom instructions. Workspace agents are team-level shared infrastructure with permissions, audit logging, durable execution, and budget controls. Custom GPTs are personal productivity tools; workspace agents are organizational platform.
How does credit-based pricing work in practice?
Each agent invocation consumes credits based on the tokens used by the underlying model, the tools the agent invokes (web search, code execution, etc.), and any external API calls. Credits are pre-purchased or auto-renewed; budgets can be set per agent and per user. A typical “research and synthesis” task consumes 50-500 credits, depending on the depth of work. Heavy data analysis tasks can consume 1,000+ credits per run.
Can workspace agents access our internal systems?
Yes, through the existing ChatGPT Enterprise connector framework. Connectors exist for SharePoint, Google Drive, Slack, Notion, Salesforce, GitHub, and dozens of other systems. Agents inherit the data access permissions configured for the workspace; they cannot access data the workspace itself can’t see.
What happens to ChatGPT Agent (the product announced earlier)?
ChatGPT Agent capabilities are now integrated into the workspace agents platform. The standalone ChatGPT Agent product converges into the broader workspace-agents framework. Customers using ChatGPT Agent will be migrated automatically; functionality is preserved.
Can we self-host workspace agents like we can with LangGraph?
No. Workspace agents are a managed OpenAI product; the runtime, state management, and tooling all live in OpenAI’s infrastructure. Customers who require self-hosted deployments should evaluate LangGraph, AutoGen, or similar self-hosted alternatives, with the trade-off being more integration work and less polish than the managed offering.
How do workspace agents integrate with our identity provider?
Through the same SSO and SCIM provisioning that ChatGPT Enterprise uses. Agents inherit user identity for permission checks; admin actions on agents (creation, editing, budget changes) require admin privileges in your identity provider. Most major IDPs (Okta, Microsoft Entra, Google Workspace, OneLogin, Auth0) integrate cleanly.