By Assaf Biton
Everyone keeps asking which company "won" the agent race. After two months tracking
OpenAI Workspace Agents, the Codex-powered shared agents built inside ChatGPT for team workflows.
Best for: General knowledge workers needing a capable all-in-one assistant, Developers wanting quick code generation with GPT Image and Codex integration
Anthropic's Claude agent stack: Managed Agents (API), Cowork (desktop), and Claude Code Agent Teams.
Best for: Knowledge workers doing deep document analysis and long-form writing, Developers who need precise instruction-following for complex multi-step tasks
TL;DR: OpenAI Workspace Agents launched April 22, 2026 in research preview, then moved to general availability on May 22, 2026 for ChatGPT Business, Enterprise, and Edu. They are GA as of June 2026, not a preview. They are Codex-powered shared agents you build inside ChatGPT and run in ChatGPT or Slack, free until July 6, 2026, then credit-based. Anthropic's comparable layer is Claude Cowork (GA April 2026, desktop app for any paid plan) plus Claude Managed Agents, the developer API at $0.08 per session-hour plus token cost. OpenAI is winning the team-adoption layer inside business tools. Anthropic still ships the deeper, better-documented runtime for engineers. Pick by which layer your problem lives in, not by brand.
The 10-Second Decision: Which Layer Is Your Problem?
Most teams overthink this. Here is the routing rule I now use, and it has nothing to do with which model benchmarks higher.
- Want non-engineers to automate repeatable work inside Slack, Gmail, or Salesforce? Start with OpenAI Workspace Agents.
- Want a developer to deploy production agents with explicit sessions, environments, and audit traces? Look at Claude Managed Agents.
- Want one person to get an autonomous desktop coworker for files and multi-step tasks? Claude Cowork.
- On ChatGPT Enterprise with EKM (customer-managed encryption keys)? Workspace Agents now work for eligible EKM workspaces, since OpenAI added EKM support on May 7, 2026. More on that below.
That EKM gap blocked regulated buyers at launch, then closed in early May. Sound familiar? Checking the current eligibility state is still the single fastest way to scope a tool before a pilot wastes a quarter.
OpenAI Workspace Agents vs Anthropic: The Quick Comparison
Here is the side-by-side I wish someone had handed me in April. Numbers are current as of June 19, 2026, pulled from each vendor's own pages.
| Dimension | OpenAI Workspace Agents | Anthropic (Cowork + Managed Agents) |
|---|---|---|
| Primary user | Business teams inside ChatGPT | Cowork: knowledge workers. Managed Agents: developers |
| Where it runs | ChatGPT and Slack, cloud execution | Cowork: macOS/Windows desktop. Managed Agents: hosted API |
| Execution substrate | Codex in the cloud (files, code, tools, memory) | Claude managed runtime plus cloud environment |
| How you build it | Conversational builder, templates, preview | Cowork: chat UI. Managed Agents: REST/SDK, sessions |
| Pricing model | Free until July 6, 2026, then credit-based | Cowork: included in paid plans. Managed Agents: $0.08/session-hour + tokens |
| Public runtime docs | Builder and admin oriented | Deep: sessions, environments, events, vaults, MCP |
| EKM support | Supported for eligible Enterprise + EKM workspaces since May 7, 2026 | Enterprise controls, but verify EKM per plan |
Read that "primary user" row twice. It explains almost every other difference. OpenAI built for the person who has never opened a terminal. Anthropic split the job in two: a desktop app for that same person, and a raw API for the engineer who wants control.
OpenAI Workspace Agents: What It Actually Is
Workspace Agents are not a renamed GPT. That was my first wrong assumption too. They are shared, Codex-powered agents that handle long-running workflows, gather context across your connected apps, ask for approval, and keep working in the cloud while you are offline.
You build one by talking to ChatGPT, pick a template, preview it, then share it privately, by link, or to a team directory. It can run on a schedule or react to a Slack message. The connector list is long: Slack, Google Drive, Gmail, Google Calendar, Microsoft apps, Salesforce, Notion, Atlassian Rovo, web search, plus custom MCP servers.
The clearest example OpenAI ships is a sales meeting prep agent. It checks your calendar, skips internal-only meetings, pulls notes from SharePoint, runs a web search, writes a 2-to-3 page brief, saves a doc, and emails a summary every morning. That is a full business workflow, not a chat reply. So the leap here is real, even if the framing is modest.
Pro Tip: Start every Workspace Agent with a dedicated service account scoped to least privilege, not your personal login. OpenAI's own docs warn that anyone the agent is shared with can act through an agent-owned connector. One scoped account up front prevents the most common data-exposure mistake in week one.
OpenAI Workspace Agents: What Works and What Doesn't
What Works
- Built inside ChatGPT, so adoption is near-zero friction for the ~900M weekly active users already there.
- Codex execution means it transforms CSVs, runs code, and renders reports, not just drafts text.
- Broad pre-wired business connectors: Slack, Salesforce, Notion, Microsoft apps, and more out of the box.
- Admin controls: role-based access, approval points, audit logs, and centralized monitoring.
What Doesn't
- EKM support is gated to eligible Enterprise workspaces and stays off until an admin enables it, so regulated buyers must confirm their workspace qualifies first.
- Credit pricing starts July 6, 2026, and a single GPT-5.5 run can burn 5 to 25 credits, so costs are hard to model until you test.
- No public low-level session or runtime API the way Anthropic documents Managed Agents.
- Shared-auth and agent-owned connectors can leak access if scoped poorly. OpenAI flags this itself.
Anthropic's Answer: Cowork, Managed Agents, and Code Teams
Here is where the comparison gets uneven, and why a head-to-head needs care. Anthropic does not have one product facing OpenAI Workspace Agents. It has three, each aimed at a slice.
Claude Cowork is the closest match for the same person. It went GA in April 2026, runs as a desktop app on macOS and Windows for any paid plan, and wraps Claude Code's agentic engine (file access, multi-step execution, tool use) in a UI anyone can drive. The April update added role-based access controls, group spend limits, and usage analytics for enterprise teams.
Claude Managed Agents is the developer layer. It is an API-first hosted runtime with Agent, Environment, Session, and Event resources, server-sent event streams, built-in tools, MCP, vaults, and outcome evaluation. It launched in public beta April 8, 2026, and it is Claude-only and cloud-only. And then there is Claude Code Agent Teams, a CLI feature for running several Claude Code sessions together. That one is experimental and off by default, so I would not build production on it yet.
Do not compare OpenAI Workspace Agents only to Claude Managed Agents. That match flatters OpenAI on ease and flatters Anthropic on depth, and misses the truth. The honest pairing is Workspace Agents against Cowork for the business user, and OpenAI Frontier against Managed Agents for the production platform.
Anthropic Claude Agents: What Works and What Doesn't
What Works
- Managed Agents has the deepest public runtime docs: sessions, environments, events, vaults, MCP, and tool policies.
- Transparent metered pricing: $0.08 per session-hour billed to the millisecond, plus standard token rates.
- Cowork gives non-developers an autonomous desktop coworker with the same engine as Claude Code.
- Permission policies and credential separation are built into the runtime, not bolted on.
What Doesn't
- The stack is split across three products, so buyers have to learn which layer solves their problem.
- Claude-only and cloud-only on Managed Agents: no multi-model fallback, no self-hosting.
- Managed Agents is developer wiring, not a pre-integrated business-app directory like ChatGPT's.
- Claude Code Agent Teams is disabled by default with caveats on resumption and coordination.
Feature by Feature: Where Each One Pulls Ahead
Benchmarks here are about packaging, not raw model IQ. Both vendors ship strong models. The difference is how the agent reaches your work.
| Capability | OpenAI Workspace Agents | Claude Managed Agents | Claude Cowork |
|---|---|---|---|
| Build without code | Yes, conversational builder | No, API/SDK only | Yes, desktop chat |
| Run in Slack | Yes, native Slack channel | Via custom integration | No, desktop only |
| Scheduled / offline runs | Yes, schedules built in | Yes, stateful sessions | Background tasks |
| Pre-wired business apps | Broadest (Salesforce, Notion, MS) | MCP servers you wire up | Connectors plus MCP |
| Low-level runtime API | Not public | Yes, fully documented | No |
| Code execution | Yes, via Codex | Yes, sandboxed environment | Yes, local file access |
The pattern is consistent. OpenAI wins the "anyone on the team can build and run this in our existing tools" race. Anthropic wins the "an engineer needs an inspectable, metered, controllable runtime" race. Both are real wins. They just serve different buyers.
The Pricing Reality Nobody Markets
This is where I burned the most time, so let me save you the hours. As of June 19, 2026, neither side is fully transparent, but they fail differently.
| Product | Cost model | Gotcha |
|---|---|---|
| OpenAI Workspace Agents | Free until July 6, 2026, then credit-based. ChatGPT Business is $20/user/month. | A single GPT-5.5 run can use 5-25 credits. Total cost is invisible until you measure real workflows. |
| Claude Managed Agents | $0.08 per session-hour, billed to the millisecond, plus standard token rates. | No batch discount on sessions. A long-running idle session still accrues runtime charges. |
| Claude Cowork | Included in existing paid Claude plans (Pro, Team, Enterprise). | Heavy agentic use can hit plan usage limits faster than chat-only use. |
So Anthropic publishes a clearer per-unit number, but session-hour billing punishes idle agents. OpenAI hides the number behind credits during the free window, then makes you reverse-engineer costs after July 6. Neither is honest enough. Both make you pilot before you can forecast.
The OpenAI free period was already pushed once, from May 6 to July 6, 2026. Do not build a 2027 budget on "it is free." Recompute your numbers the first week of July, or you could be off by several multiples.
The Governance Trap Both Sides Share
Here is the thing about agents that act on your behalf: a misconfigured connector is a data leak waiting to happen. To be fair, both vendors document this honestly rather than hiding it.
OpenAI's help center spells out the risks plainly. Write-action safety, shared Slack auth where others can act through your connection, agent-owned accounts that expose resources to anyone the agent is shared with, and personal credentials leaking through role-based publishing. The fix is the same every time: scoped service accounts, least privilege, approval gates before any write, and audit review.
Anthropic builds permission policies, credential separation, and tracing into the Managed Agents runtime itself. That is a structural advantage for engineers. But Cowork still hands a desktop agent real file and tool access, so the same discipline applies. Neither product is safe by default. They are safe when configured by someone who read the warnings.
Pro Tip: Gate every write action behind human approval for the first 30 days of any pilot, on either platform. You will catch roughly the scary mistakes (wrong recipient, over-broad query, accidental delete) before they touch production data. Turn approvals off only per-workflow, once you trust it.
Three Reader Scenarios. Pick Yours.
The Ops Lead Who Wants Slack Automations Today
You run revenue ops or support. You want weekly briefs, ticket triage, and meeting prep, and your team lives in Slack and Salesforce. Honest recommendation: pilot OpenAI Workspace Agents now, while it is free until July 6. The connector breadth and Slack-native runs fit your work better than anything Anthropic ships at the team layer. Just start with a service account and draft-only outputs.
The Engineer Building a Production Agent
You need an agent in your own product with inspectable sessions, metered cost, and audit traces. Go with Claude Managed Agents. The $0.08 per session-hour model is the most transparent unit price in this matrix, and the public session/environment/event API gives you control Workspace Agents simply does not expose yet. Watch idle sessions, since runtime billing does not stop while they sit.
The Regulated Enterprise on ChatGPT Enterprise + EKM
Healthcare, finance, public sector. You already standardized on ChatGPT Enterprise with customer-managed keys. Workspace Agents are now available to eligible EKM workspaces, since OpenAI added EKM support on May 7, 2026. Confirm your workspace qualifies and have an admin enable it, then pilot it alongside Claude Cowork or Managed Agents with your security team. The launch-day blocker is gone, so judge it on governance fit, not availability.
Frequently Asked Questions
Are OpenAI Workspace Agents available now?
Yes. They launched in research preview on April 22, 2026, then reached general availability on May 22, 2026 for ChatGPT Business, Enterprise, and Edu. Business has them enabled by default. Enterprise and Edu require an admin to turn them on, and eligible Enterprise workspaces using EKM can now create and use them too after OpenAI added EKM support on May 7, 2026.
How much do OpenAI Workspace Agents cost?
They are free until July 6, 2026 (a date already extended once from May 6), then credit-based. A typical end-to-end run on GPT-5.5 uses roughly 5 to 25 credits depending on the workflow. ChatGPT Business itself is $20 per user per month. Enterprise and Edu pricing is custom.
Is Anthropic better than OpenAI for AI agents?
Neither is "better" across the board. Anthropic Claude Managed Agents has deeper public runtime docs and clearer per-session pricing, which engineers prefer. OpenAI Workspace Agents has broader pre-wired business connectors and lives inside ChatGPT and Slack, which non-developers adopt faster. Match the tool to your buyer, not to a brand.
What is the difference between Workspace Agents and OpenAI Frontier?
Workspace Agents are the team layer inside ChatGPT for repeatable business workflows. OpenAI Frontier is the deeper enterprise platform for production AI coworkers, with systems-of-record integration, durable memory, agent execution, evaluation, and governance. Together they form OpenAI's two-layer agent strategy. See our comparisons hub for the rest of the field.
Can OpenAI Workspace Agents replace a developer's coding agent?
Not yet. They are a ChatGPT/Slack business-workflow layer, not a local engineering control plane. For parallel coding work, Claude Code, Codex CLI, or Claude Code Agent Teams still fit better. Workspace Agents shine on reports, triage, and meeting prep, not on navigating a large repo.
Our Recommendation
Best for team adoption inside business tools: OpenAI Workspace Agents. The ChatGPT-native builder plus Slack and Salesforce connectors get non-developers shipping useful automations fast.
Best for a controllable production runtime: Claude Managed Agents. The documented session/environment API and $0.08 per session-hour pricing give engineers control and a clear unit cost.
Best autonomous desktop coworker: Claude Cowork. Same engine as Claude Code, wrapped in a UI any paid-plan user can drive.
Watch the fine print: OpenAI Workspace Agents on ChatGPT Enterprise with EKM. Support arrived May 7, 2026, but it is limited to eligible workspaces and stays off until an admin turns it on, so confirm your eligibility before you plan around it.
Want to go deeper? Read our agent permissions playbook before you wire up any connector, then check the 2026 agent operating model for the org-level picture. When you are ready to pilot, start one low-risk workflow with draft-only output this week, measure the credits, and decide on real numbers instead of a launch blog.
