{"title":"Agent Harness Comparison","source":"https://resources.ascendlabs.ai/openai-vs-anthropic-for-teams","prepared_by":"Prepared by AscendAI","last_verified":"2026-07-10","columns":[{"id":"oaBiz","vendor":"OpenAI","label":"ChatGPT Business"},{"id":"oaEnt","vendor":"OpenAI","label":"ChatGPT Enterprise"},{"id":"clTeam","vendor":"Anthropic","label":"Claude Team"},{"id":"clEnt","vendor":"Anthropic","label":"Claude Enterprise"}],"matrix":[{"id":"sso","cat":"Identity and access","claim":"Single sign-on (SSO / SAML)","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"yes","t":"Yes"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"scim-provisioning","cat":"Identity and access","claim":"SCIM provisioning from identity provider","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"No"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"domain-verification","cat":"Identity and access","claim":"Domain verification / domain capture","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"No"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"custom-roles-rbac","cat":"Identity and access","claim":"Custom roles and user groups (RBAC)","values":{"oaBiz":{"s":"no","t":"Owner / admin / member only"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"Admin / user only"},"clEnt":{"s":"yes","t":"Groups, manual or SCIM"}},"detail":"Both vendors gate real role granularity behind Enterprise. Claude Enterprise groups can be managed manually or synced from your identity provider via SCIM.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"ip-allowlists","cat":"Identity and access","claim":"IP allowlists","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Not offered"}},"detail":"Anthropic does not offer IP allowlisting at any tier. If network-level access control is a hard requirement, this is an OpenAI Enterprise capability only.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"shared-team-agents","cat":"Agent capabilities","claim":"Shared team agents (build once, whole org runs them)","values":{"oaBiz":{"s":"yes","t":"Workspace Agents"},"oaEnt":{"s":"yes","t":"Workspace Agents"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Plugins / skills only"}},"detail":"This is OpenAI’s clearest structural lead: one person builds an agent, publishes it to a workspace directory, and anyone runs it. Anthropic distributes know-how through admin-provisioned plugins and skills; there are no shared running agents.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"personal-agentic-sessions","cat":"Agent capabilities","claim":"Personal agentic sessions","values":{"oaBiz":{"s":"yes","t":"Agent mode"},"oaEnt":{"s":"yes","t":"Agent mode"},"clTeam":{"s":"yes","t":"Cowork"},"clEnt":{"s":"yes","t":"Cowork"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"agent-access-granularity","cat":"Agent capabilities","claim":"Agent access control granularity","values":{"oaBiz":{"s":"partial","t":"Feature toggle"},"oaEnt":{"s":"yes","t":"Per role, per app, per action"},"clTeam":{"s":"partial","t":"Org-wide all-or-nothing toggle"},"clEnt":{"s":"yes","t":"Per group via custom roles"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"cloud-execution-devices-off","cat":"Agent capabilities","claim":"Cloud execution with all devices off","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"partial","t":"Beta, rolling out"},"clEnt":{"s":"partial","t":"Beta, rolling out"}},"detail":"Anthropic’s historical weakness, agents tethered to a running desktop, was largely eliminated on July 7, 2026 when Cowork sessions moved to cloud execution by default. Still in beta, rolling out by plan.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-07"},{"id":"scheduled-autonomous-tasks","cat":"Agent capabilities","claim":"Scheduled autonomous tasks","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"yes","t":"Cloud, device-independent"},"clEnt":{"s":"yes","t":"Cloud, device-independent"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"local-file-manipulation","cat":"Agent capabilities","claim":"Local file / desktop app manipulation","values":{"oaBiz":{"s":"no","t":"Browser sandbox only"},"oaEnt":{"s":"no","t":"Browser sandbox only"},"clTeam":{"s":"yes","t":"Via desktop app"},"clEnt":{"s":"yes","t":"Via desktop app"}},"detail":"Claude’s desktop app can work on local files and local applications directly. ChatGPT agents operate inside a sandboxed virtual browser and never touch the local machine.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"coding-agent-standard-seat","cat":"Agent capabilities","claim":"Coding agent included in standard seat","values":{"oaBiz":{"s":"no","t":"Codex seats closed to new Business workspaces since 6/24"},"oaEnt":{"s":"yes","t":"Codex, usage-billed"},"clTeam":{"s":"yes","t":"Claude Code, every seat"},"clEnt":{"s":"yes","t":"Claude Code, every seat"}},"detail":"Since June 24, 2026, new ChatGPT Business workspaces cannot add Codex seats; coding-agent access is effectively Enterprise-gated for new OpenAI customers. Claude Code ships in every seat on both Anthropic tiers.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-06-24"},{"id":"non-overridable-policies","cat":"Agent capabilities","claim":"Non-overridable managed agent policies","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"partial","t":"Per-action safeguards and confirmation flows"},"clTeam":{"s":"yes","t":"Managed settings, MDM keys"},"clEnt":{"s":"yes","t":"Managed settings, MDM keys"}},"detail":"Anthropic’s managed policies are enforced at the device level via MDM keys; users cannot override them. OpenAI Enterprise offers per-action safeguards and confirmation flows rather than hard device-level enforcement.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"per-connector-restrictions","cat":"Agent capabilities","claim":"Per-connector read/write restrictions","values":{"oaBiz":{"s":"partial","t":"App enable / disable"},"oaEnt":{"s":"yes","t":"Per action"},"clTeam":{"s":"partial","t":"Org-wide, per tool"},"clEnt":{"s":"yes","t":"Per group, per tool"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"shared-projects","cat":"Collaboration and output sharing","claim":"Shared Projects","values":{"oaBiz":{"s":"yes","t":"Collaborative conversations"},"oaEnt":{"s":"yes","t":"Collaborative conversations"},"clTeam":{"s":"partial","t":"Shared knowledge only; chats stay private"},"clEnt":{"s":"partial","t":"Shared knowledge only; chats stay private"}},"detail":"In ChatGPT, teammates work inside the same conversations with shared files, instructions, and project-scoped memory. In Claude, a shared project is a reference library; every chat inside it stays private to its author unless shared as a snapshot.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"realtime-co-editing","cat":"Collaboration and output sharing","claim":"Real-time co-editing of outputs","values":{"oaBiz":{"s":"yes","t":"Canvas: cursors, comments, versions"},"oaEnt":{"s":"yes","t":"Canvas: cursors, comments, versions"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Not offered"}},"detail":"Canvas co-editing (live cursors, inline comments, version history, section assignments) reached all ChatGPT plans in July 2026. Claude has no multi-user conversation and no real-time co-editing.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-01"},{"id":"publish-interactive-apps","cat":"Collaboration and output sharing","claim":"Publish outputs as interactive apps / portals","values":{"oaBiz":{"s":"yes","t":"Sites"},"oaEnt":{"s":"yes","t":"Sites"},"clTeam":{"s":"partial","t":"Artifacts, org-internal only"},"clEnt":{"s":"partial","t":"Artifacts, org-internal only"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"public-url-sharing","cat":"Collaboration and output sharing","claim":"Public URL sharing of published outputs","values":{"oaBiz":{"s":"yes","t":"Yes (July 2026)"},"oaEnt":{"s":"yes","t":"Yes (July 2026)"},"clTeam":{"s":"no","t":"Free / Pro / Max only"},"clEnt":{"s":"no","t":"Free / Pro / Max only"}},"detail":"A red flag that cuts across tiers: Claude org plans cannot publish outputs to public URLs at any tier. For client-facing work, outputs must be exported as files or shared behind org authentication.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-01"},{"id":"office-add-ins","cat":"Collaboration and output sharing","claim":"First-party Microsoft Office add-ins","values":{"oaBiz":{"s":"partial","t":"Excel (GA), PowerPoint (beta) only"},"oaEnt":{"s":"partial","t":"Excel (GA), PowerPoint (beta) only"},"clTeam":{"s":"yes","t":"Excel, Word, PowerPoint (GA) + Outlook (beta)"},"clEnt":{"s":"yes","t":"Excel, Word, PowerPoint (GA) + Outlook (beta)"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"cross-app-context","cat":"Collaboration and output sharing","claim":"Cross-app context carry between Office apps","values":{"oaBiz":{"s":"no","t":"Add-in chats siloed from ChatGPT history"},"oaEnt":{"s":"no","t":"Add-in chats siloed from ChatGPT history"},"clTeam":{"s":"yes","t":"One thread across all four apps"},"clEnt":{"s":"partial","t":"Default OFF; org owner must enable"}},"detail":"On Claude, one conversation thread follows the work across Outlook, Word, Excel, and PowerPoint. On Claude Enterprise this is off by default until an org owner enables it. ChatGPT’s add-in sessions are siloed by design.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"ai-design-workspace","cat":"Built-in add-ons","claim":"AI design workspace","values":{"oaBiz":{"s":"no","t":"Image gen + Canva app; no dedicated tool"},"oaEnt":{"s":"no","t":"Image gen + Canva app; no dedicated tool"},"clTeam":{"s":"yes","t":"Claude Design (research preview)"},"clEnt":{"s":"partial","t":"Claude Design; disabled by default, admin enables"}},"detail":"Claude Design (research preview, April 2026) turns prompts, docs, or codebases into interactive prototypes, decks, one-pagers, and landing pages, and builds a design system from your codebase and brand files so output stays on-brand. Exports to PPTX, PDF, HTML, URL, or Canva.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"native-image-generation","cat":"Built-in add-ons","claim":"Native image generation","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Not offered"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"voice-mode","cat":"Built-in add-ons","claim":"Voice mode","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Not offered"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"custom-agent-app-store","cat":"Built-in add-ons","claim":"Custom agent / app store","values":{"oaBiz":{"s":"yes","t":"GPT Store (millions of GPTs) + Apps SDK"},"oaEnt":{"s":"yes","t":"GPT Store + workspace-internal GPTs"},"clTeam":{"s":"partial","t":"Plugin directory (smaller)"},"clEnt":{"s":"yes","t":"Plugins + private Git-based marketplaces"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"browser-agent","cat":"Built-in add-ons","claim":"Browser agent","values":{"oaBiz":{"s":"yes","t":"Agent mode virtual browser (Atlas retiring 8/9)"},"oaEnt":{"s":"yes","t":"Agent mode virtual browser"},"clTeam":{"s":"yes","t":"Claude in Chrome (user’s real browser)"},"clEnt":{"s":"yes","t":"Claude in Chrome (admin can disable)"}},"detail":"OpenAI’s standalone Atlas browser retires August 9, 2026 as agentic browsing folds into ChatGPT and Codex. Claude in Chrome operates the user’s actual browser with their logged-in sessions; powerful, but a different risk profile.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-08-09"},{"id":"slack-coworker-agent","cat":"Built-in add-ons","claim":"Slack coworker agent","values":{"oaBiz":{"s":"yes","t":"Workspace Agents run in Slack"},"oaEnt":{"s":"yes","t":"Workspace Agents run in Slack"},"clTeam":{"s":"yes","t":"Claude Tag"},"clEnt":{"s":"yes","t":"Claude Tag"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"usage-pooling","cat":"Spend and usage","claim":"Usage pooling across the team","values":{"oaBiz":{"s":"yes","t":"Workspace credit pool"},"oaEnt":{"s":"yes","t":"Workspace credit pool"},"clTeam":{"s":"no","t":"Per-seat limits"},"clEnt":{"s":"no","t":"Per-seat limits"}},"detail":"The structural difference in economics: OpenAI’s pooled credits absorb spiky usage across a team. Claude’s per-seat limits mean your two or three power users hit individual ceilings even while the rest of the team has headroom.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"org-user-spend-caps","cat":"Spend and usage","claim":"Org-level and per-user spend caps","values":{"oaBiz":{"s":"partial","t":"Plan credit pool"},"oaEnt":{"s":"yes","t":"Budgets / limits per chat or project"},"clTeam":{"s":"yes","t":"Yes"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"per-group-spend-limits","cat":"Spend and usage","claim":"Per-group spend limits","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"partial","t":"Per chat / project, not per group"},"clTeam":{"s":"no","t":"No"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"usage-analytics-dashboard","cat":"Spend and usage","claim":"Usage analytics dashboard","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"yes","t":"Yes"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"analytics-api","cat":"Spend and usage","claim":"Analytics API","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"yes","t":"Yes"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"full-audit-logs","cat":"Oversight and compliance","claim":"Full audit logs","values":{"oaBiz":{"s":"partial","t":"Usage analytics only"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"No"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"compliance-api","cat":"Oversight and compliance","claim":"Compliance API (programmatic content access)","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"No"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"agent-activity-audit-trail","cat":"Oversight and compliance","claim":"Agent activity captured in audit trail","values":{"oaBiz":{"s":"partial","t":"Admin console view of agent runs"},"oaEnt":{"s":"yes","t":"Agent activity + audit / export"},"clTeam":{"s":"no","t":"Cowork not in audit logs"},"clEnt":{"s":"no","t":"Cowork not in audit logs"}},"detail":"A red flag that cuts across tiers: Cowork activity does not appear in Claude’s audit trail at any tier. The same gap currently applies to Claude’s Office add-ins. If audit coverage of agent work is a requirement, this is the row to watch.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"otel-siem-streaming","cat":"Oversight and compliance","claim":"OpenTelemetry / SIEM event streaming","values":{"oaBiz":{"s":"no","t":"Not offered"},"oaEnt":{"s":"no","t":"Not offered"},"clTeam":{"s":"yes","t":"Yes"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"enterprise-key-management","cat":"Oversight and compliance","claim":"Enterprise Key Management","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Not offered"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"multi-region-residency","cat":"Oversight and compliance","claim":"Multi-region data residency","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"no","t":"Not offered"},"clEnt":{"s":"no","t":"Not offered"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"configurable-retention","cat":"Oversight and compliance","claim":"Configurable retention / Zero Data Retention","values":{"oaBiz":{"s":"no","t":"No"},"oaEnt":{"s":"partial","t":"Retention policies"},"clTeam":{"s":"no","t":"No"},"clEnt":{"s":"yes","t":"Yes, incl. ZDR"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"id":"no-training-on-data","cat":"Oversight and compliance","claim":"No training on organizational data","values":{"oaBiz":{"s":"yes","t":"Yes"},"oaEnt":{"s":"yes","t":"Yes"},"clTeam":{"s":"yes","t":"Yes"},"clEnt":{"s":"yes","t":"Yes"}},"detail":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"}],"dimensions":[{"id":"credits","intro":"","rows":[{"d":"Base allocation","oa":"Each seat includes baseline usage of chat and core features.","cl":"Each seat includes rolling 5-hour and weekly usage limits per individual user, shared across Chat, Claude Code, and Cowork. Limits doubled May 6, 2026.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Pooled or individual?","oa":"Pooled. Agentic features draw from a shared workspace credit pool after included usage. One heavy user can draw on the team’s pool.","cl":"Individual. Limits are per seat. One user’s unused capacity cannot transfer to another. Heavy users hit their own ceilings regardless of team headroom.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Overage mechanics","oa":"Token-based credit pricing for agent runs (input, cached input, output tokens) effective July 6, 2026. Admins buy workspace credits and view breakdown by user, product, and model.","cl":"Admin-enabled extra usage billed at standard API rates, with admin-set per-user spend caps. Enterprise adds per-group spend limits.","changeNote":"OpenAI moved agent-run overages to token-based workspace credit pricing, effective July 6, 2026.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-06"},{"d":"Heavy-user path","oa":"Credit purchases from the workspace pool.","cl":"Premium seats ($100/seat/mo on Team) for users needing significantly more.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Admin visibility","oa":"Global Admin Console consolidates credit usage by user, product, and model.","cl":"Dashboard plus Analytics API: per-user activity, skill and connector invocations, per-repo Claude Code usage.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"}],"callout":{"title":"Working assumption on burn rate","body":"Assume Claude consumes usage allocation materially faster than ChatGPT for equivalent work. Anthropic’s own documentation notes Cowork tasks consume significantly more usage than chat. This may converge, so budget for it now. For a 10-person team with 2 heavy agentic users, plan on premium seats or extra-usage budget on Claude from day one; on ChatGPT, plan a workspace credit budget and set spend alerts before enabling Workspace Agents."}},{"id":"interaction","intro":"This is the largest day-to-day operating difference between the platforms, and it favors OpenAI more than the technical comparisons suggest.","rows":[{"d":"Daily surface","oa":"One browser-based workspace. Nothing to install; works on any OS including Linux and Chromebooks.","cl":"Claude app on web, desktop, and mobile. Cowork’s full local capabilities require the desktop app on macOS or Windows.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"What a shared Project is","oa":"A collaborative container: teammates work inside the same conversations, with shared files, shared instructions, and project-scoped memory the whole team builds on.","cl":"A shared reference library: teammates can start their own chats grounded in its knowledge, but every chat inside stays private to its author unless shared as a snapshot.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Context permeability","oa":"Project memory is siloed from main chat by design; within a shared project, context compounds across the team.","cl":"Same siloing, plus a second wall: work done inside a project largely stays with the individual who did it.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Co-editing outputs","oa":"Canvas: live cursors, inline comments, version history, section assignments, on all plans as of July 2026.","cl":"No multi-user conversation and no real-time co-editing of outputs.","changeNote":"Canvas real-time co-editing reached all ChatGPT plans in July 2026.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-01"},{"d":"Distributing team know-how","oa":"Workspace Agents directory: one person builds an agent, publishes it, anyone runs it.","cl":"Plugins and skills provisioned by admins; an activity feed for sharing chat snapshots. No shared running agents.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Producing deliverables","oa":"Sites turns work into dashboards, trackers, portals, prototypes, and reports inside ChatGPT. Excel and PowerPoint sidebars assist in Office.","cl":"Cowork produces real files: Excel with working formulas, PowerPoint, Word. First-party add-ins for all four Office apps write in place with tracked changes. Claude Design produces prototypes, decks, and one-pagers.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Handing outputs to a client","oa":"Publish a Site publicly and send the URL (Business and Enterprise, July 2026).","cl":"No public artifact URLs on Team / Enterprise. Export the file or grant org-authenticated access. Claude Code Artifacts publish to private org URLs only.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"}],"note":"Net: for teams building compounding shared context, Claude’s Projects remain its most frustrating structural limitation, and the absence of public publishing on org plans compounds it for client-facing work. Claude’s counterweight: its outputs are real files in your real applications, not pages living inside the vendor’s walls."},{"id":"office","intro":"A genuine Anthropic advantage, and the asymmetry is larger than most comparisons acknowledge. Ironically, given OpenAI’s Microsoft partnership, Anthropic ships the more complete first-party Office integration; Microsoft reserves the deep native slot for Copilot.","rows":[{"d":"App coverage (first-party add-ins)","oa":"Excel (GA) and PowerPoint (beta) only, via one AppSource listing. No first-party Word or Outlook add-in.","cl":"All four: Excel, Word, and PowerPoint generally available (May 7, 2026), Outlook in public beta. Deployable by admins from the Microsoft admin center.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Cross-app handoff","oa":"None. Add-in chats operate separately from ChatGPT.com conversations, do not sync with chat history, and Memory is unavailable in the add-ins.","cl":"One conversation thread carries across all four apps: triage an email in Outlook, draft the memo in Word, build the model in Excel, generate the deck in PowerPoint, without re-explaining.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Where context is stored","oa":"Add-in sessions are siloed by design; connector reads (OneDrive, SharePoint, Outlook, Teams) pull data into the main ChatGPT context under workspace policies.","cl":"Cross-app context lives in the conversation thread, processed on Anthropic’s backend and deleted within 30 days. Claude only sees files that are currently open.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Admin controls and defaults","oa":"Admins enable the add-in and connectors per workspace; connector write actions are admin-gated and off by default.","cl":"Cross-app work is default OFF for Team / Enterprise; an org owner must enable it. OpenTelemetry can stream prompts, tool calls, and document references to your collector.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Governance caveat","oa":"Add-in activity governed under standard workspace compliance.","cl":"The Office add-ins do NOT inherit org custom retention settings, and add-in activity is not currently in Enterprise audit logs, the Compliance API, or data exports. Same gap pattern as Cowork.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Data-layer connector (separate from add-ins)","oa":"OneDrive, SharePoint, Outlook, and Teams connectors; read plus admin-gated write actions since March 2026.","cl":"M365 MCP connector reads Outlook, OneDrive, SharePoint, and Teams from Cowork sessions. Read-only, user-delegated permissions, data not cached beyond the session.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"}],"note":"Operating summary: on Claude, one working session follows the work across email, memo, model, and deck. On ChatGPT, each Office app is a separate assistant and consolidation happens in the ChatGPT window via connectors. For regulated clients, note that both of Claude’s most differentiated surfaces (Cowork and the Office add-ins) currently sit outside its audit and retention framework."},{"id":"addons","intro":"","rows":[{"d":"Extension architecture","oa":"Apps SDK and connectors (60+ first-party); apps in ChatGPT from consumer brands; Agent Builder and ChatKit for custom agents.","cl":"Model Context Protocol (MCP), an open standard: connector directory plus any custom MCP server. Skills (portable procedure files) and plugins bundle skills, connectors, slash commands, and sub-agents.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Marketplace scale","oa":"GPT Store with millions of custom GPTs. For niche tasks a GPT usually already exists. Workspace-internal GPT sharing for orgs.","cl":"Plugin directory is much smaller (11+ official role plugins: Sales, Legal, Marketing, Finance). Strength is custom internal plugins, not breadth.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Private distribution","oa":"Internal-only GPTs shared within the workspace; admin RBAC over who creates and publishes.","cl":"Private plugin marketplaces synced from a Git repository, assignable to directory groups; vertical agent template packs.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Design tooling","oa":"No dedicated design product; native image generation plus a Canva app cover some ground.","cl":"Claude Design (research preview): prompts, docs, or codebases become prototypes, decks, one-pagers, and landing pages, on-brand, exportable to PPTX, PDF, HTML, URL, or Canva. Enterprise: disabled by default, admin must enable.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Native media","oa":"Image generation, voice mode, and video (Sora) built in.","cl":"No image generation and no voice mode in the core product; visual work routes through Claude Design or external tools.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Browser presence","oa":"Agent mode’s virtual browser handles web tasks in a sandbox; the standalone Atlas browser retires August 9, 2026.","cl":"Claude in Chrome operates the user’s actual browser with their logged-in sessions; powerful, but a different risk profile, and admin-disableable.","changeNote":"OpenAI’s standalone Atlas browser is scheduled to retire August 9, 2026 as agentic browsing folds into ChatGPT and Codex.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-08-09"},{"d":"Coding ecosystem","oa":"Codex app with plugin sharing (admin-gated); DigitalOcean droplet workspaces. Codex seats now Enterprise-only for new orgs.","cl":"Claude Code in every seat with skills, hooks, MCP servers, subagents, and a plugin ecosystem. Claude Code Artifacts publish session results as live pages (Team / Enterprise beta).","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"}],"note":"Pattern: OpenAI’s ecosystem is broad and consumer-scale: millions of GPTs, native media, brand apps. Anthropic’s is narrower but deeper and more governable: open MCP standard, Git-based private marketplaces, design-system-aware tooling. For an agency building client-specific capability, Claude’s plugin and skills model is the stronger raw material; for end users who want something off the shelf today, the GPT Store usually has it."},{"id":"architecture","intro":"","rows":[{"d":"Primary agent surfaces","oa":"Workspace Agents (cloud, shared), Agent mode / Operator (sandboxed browser VM), Codex (cloud + local host machine).","cl":"Cowork (cloud sessions by default as of July 7, 2026; local desktop execution still available), Claude Code (local + cloud web), Claude in Chrome.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Runs with all devices off?","oa":"Yes. Workspace agents are cloud-hosted; schedules fire with no one in ChatGPT.","cl":"Yes, as of July 2026. Cowork sessions and scheduled tasks run on Anthropic’s servers; no device needs to be awake.","changeNote":"Anthropic moved Cowork sessions to cloud execution by default on July 7, 2026, so agents run with all devices off.","source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-07"},{"d":"Remaining local dependency","oa":"Codex local work requires an awake host Mac / Windows machine; the mobile app steers it remotely.","cl":"Local file access, local connectors, browser use, and computer use require the desktop app open; sessions otherwise continue in the cloud.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Sandboxing","oa":"Agent mode runs in a virtual browser / VM; locked computer-use controls for admins.","cl":"Per-session isolated sandboxes, destroyed at session end; mandatory egress proxy with allow-listed destinations; connector tokens never enter the sandbox.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"},{"d":"Platform reach","oa":"Fully browser-based; Linux and Chromebook friendly.","cl":"Desktop app is macOS / Windows only; computer use unavailable on Linux; web / mobile Cowork closes most of the gap.","changeNote":null,"source_url":null,"last_verified":"2026-07-10","last_changed":"2026-07-10"}]}],"changelog":[{"date":"2026-08-09","kind":"scheduled","anchor":"sec-addons","section":"The add-on universe and built-in extras","subject":"Browser presence","detail":"OpenAI’s standalone Atlas browser is scheduled to retire August 9, 2026 as agentic browsing folds into ChatGPT and Codex."},{"date":"2026-07-07","kind":"changed","anchor":"sec-execution","section":"Agent execution architecture","subject":"Runs with all devices off?","detail":"Anthropic moved Cowork sessions to cloud execution by default on July 7, 2026, so agents run with all devices off."},{"date":"2026-07-06","kind":"changed","anchor":"sec-credits","section":"Credits and usage economics: pooled vs. individual","subject":"Overage mechanics","detail":"OpenAI moved agent-run overages to token-based workspace credit pricing, effective July 6, 2026."},{"date":"2026-07-01","kind":"changed","anchor":"sec-matrix","section":"Capability matrix: all four plans, side by side","subject":"Public URL sharing of published outputs","detail":"A red flag that cuts across tiers: Claude org plans cannot publish outputs to public URLs at any tier. For client-facing work, outputs must be exported as files or shared behind org authentication."},{"date":"2026-07-01","kind":"changed","anchor":"sec-team","section":"Team interaction: what each platform looks like to a user","subject":"Co-editing outputs","detail":"Canvas real-time co-editing reached all ChatGPT plans in July 2026."},{"date":"2026-06-24","kind":"changed","anchor":"sec-matrix","section":"Capability matrix: all four plans, side by side","subject":"Coding agent included in standard seat","detail":"Since June 24, 2026, new ChatGPT Business workspaces cannot add Codex seats; coding-agent access is effectively Enterprise-gated for new OpenAI customers. Claude Code ships in every seat on both Anthropic tiers."}]}