Claude (including Cowork execution)
Anthropic
As of 9 Oct 2026, PAEval has compiled 23 capability records for Claude (including Cowork execution) from 19 source URLs; 1 are independent observations and 19 come from the vendor's own documentation. No independent product-level score has been published for it yet. Scores from different evaluators use different scales and are not combined. PAEval does not publish its own score or rank for it.
On 2026-09-16, it was announced that chat and Cowork were merged into Claude, and Pro/Max was gradually launched; Team/Enterprise still needs to distinguish between the old selector and organization switch. Claude Code is not incorporated into the bank's ability score.
No independent product-level score has been published for this product yet. Capability evidence below is documentation and observations only.
- Independent1
- Vendor19
- Third-party report0
- Not established3
23 records · 19 source URLs · 1 external result(s)
Capability summary
Strongest status per domain; every underlying record is listed below. Availability does not establish task success.
Positioning & scope
On 2026-09-16, it was announced that chat and Cowork were merged into Claude, and Pro/Max was gradually launched; Team/Enterprise still needs to distinguish between the old selector and organization switch. Claude Code is not incorporated into the bank's ability score.
Conditions: Fixed package, whether the account has been migrated, platform and observation date
Limitations: Not all users have completed the merge.
Regions, eligibility & access
Web, macOS/Windows desktop, iOS/Android can be accessed. The official region whitelist includes the United States, the United Kingdom, Taiwan, Singapore, etc.; Mainland China and Hong Kong are not included and are not supported regions according to this policy.
Limitations: Regional policies also restrict the registration place and control relationship of some entities; access cannot be changed by changing the language.
There are 11 official web/desktop interface languages listed; instructions can be used for conversations in other languages. Chinese dialogue, Chinese interface and voice support cannot be equated.
Limitations: No cross-language execution quality testing has been done.
Pricing & quotas
Free $0; Pro $20/month or $200/year; Max starts from $100/month, 5x/20x usage level. Prices do not include applicable taxes, credits will be consumed by tasks/codes/tools.
Limitations: The specific price of Max 20x has not been verified separately from the pricing text this time; it is not an unlimited mission.
Team standard seat is $25/month or annual discount is $20/seat/month; Premium is $125/month or annual discount is $100/seat/month; Enterprise is $20/seat/month and annual fee is charged according to the API rate.
Conditions: The Team page is marked with 2-150 people; the quotation based on the account shall prevail.
Autonomy & background work
All paid plans support scheduled tasks; new cloud tasks can be run even after shutdown, and can be paused/resumed/manually run/viewed past results. Tasks that require local resources are still subject to device online restrictions.
Limitations: Any connector event triggers, autonomous discovery without settings and continuous active service are not verified this time; it is not written as not supported.
This time to confirm the scheduled task, personal Claude's native trigger matrix for any connection service event has not been confirmed.
Execution environments
Starting from October 6th, new Pro/Max tasks will be in the cloud and cannot be switched back to pure local; previous local tasks will be retained. Access to local files, MCP, browser and desktop controls still relies on the desktop application being online.
Limitations: The continuous operation of the cloud does not mean that the offline computer can still be operated; the mobile phone is a remote entrance and does not prove native mobile UI control.
Pro/Max's macOS/Windows computer control is beta; prioritize the connector, then the browser, then the screen. macOS 15+ can work in the background window; per-app authorization, block list can be set.
Limitations: Screen control directly affects the real desktop, without sandboxing at the same level; this function is currently not available in Team/Enterprise.
Calls, messages & identity
The Gmail connector can send, reply, and forward from the user's authorized mailbox; @Claude is the Slack robot entrance for Team/Enterprise beta.
Limitations: Unverified personal Claude will automatically assign a separate email address, mobile phone number or payment wallet.
Connectors & tools
The Google connector can send emails, manage calendars, and share/move/delete Drive files; the Docs/Sheets/Slides connector beta can be created and edited on October 6. The desktop supports local MCP extensions, and cloud sessions access local MCP through the online desktop.
Conditions: Organization switches, Google permissions and per-tool authorization.
Limitations: The content of Gmail attachments cannot be read directly by this connector, only metadata; the Anthropic model API is not treated as a personal Claude full agent API.
There is no public API that can be used programmatically to call the entire personal Claude execution experience.
Limitations: The model API, Claude Code interface and the complete personal agent product are not the same entity.
Memory & personalization
Cloud Cowork and chat share theme memory; projects have separate memories. Can view/edit/delete, pause/reset, close before starting a single session; can import and export. Deleting the original chat does not automatically delete the derived memory.
Conditions: Free/Pro/Max are enabled by default; Team/Enterprise requires organization permission and member activation.
Limitations: The old local Cowork does not have this shared memory; sensitive topics are not saved by default and can be selected individually.
Channels & interfaces
The unified conversation can move from Q&A to task execution; tasks can be viewed and guided across web/desktop/mobile phones; the Chrome sidebar is also the entrance. The voice beta covers mobile/web/desktop; the real-time help page clarifies that the new version of the unified experience can speak during the conversation of performing tasks, and the old Cowork/Code only supports dictation.
Research & deliverables
Supports online retrieval, paid research, code execution and xlsx/pptx/docx/PDF production; authorized files and tools can be used for tasks.
Limitations: The functional description is not the output accuracy; Claude Code’s exclusive development capabilities are recorded separately.
Docs/Slides/Design has exited beta and covers Free; Dashboards is in paid beta, and Motion is in Team/Enterprise beta. Motion uses editable code to generate animations and export MP4, not video generation models.
Conditions: The default activation time of Enterprise Docs/Slides/Design is October 15th, and can be enabled in advance by the administrator on October 9th.
Transactions & bookings
The calendar can create/change/delete events and invitations, and check shared availability. Browsers/desktops can perform cross-site tasks, but do not promise universal ordering or closed payment loops; officials recommend not using computers to control sensitive tasks such as banking and health.
Limitations: The specific merchants/regions/success rates for product payment and restaurant reservations are unknown; this cannot be the same as rejecting all consumption assistance.
Collaboration & parallelism
Announced on October 8, the support team can jointly edit documents, slides, designs, and Kanban boards with Claude, and can share external links with the administrator's permission. @Claude Collaboration is possible in threads.
Limitations: The hard upper limit for the concurrency of personal background tasks and the conflict resolution of instructions across multiple real users have not been verified.
The hard limit for individual execution tasks to run simultaneously has not been verified.
Reliability & recovery
Cloud session isolation; Team/Enterprise can use OpenTelemetry, and Compliance API can capture Cowork activities; file/web page tools still run when the local VM is unavailable, and code execution reports workspace unavailable. Endpoint EDR cannot observe activities within the sandbox or in the cloud.
Limitations: There are no independently measured success rates/SLAs at the product level.
2026-09-08 User reported that Windows local scheduled task permissions have returned, providing version, log and reproduction steps; the issue is still Open when opened. It is only a first-hand failure report of a specific version and does not prove that the new version of the cloud task in October also failed.
Permissions, security & data
Manual approval, automatic security review and skipped approval cannot be confused; permanent file deletion officially requires explicit permission. PC controls are not subject to the same tool-by-tool checks as other tools and require application-level permissions; the manufacturer acknowledges that injection protection is not absolute.
Local files read by cloud tasks are also processed on the Anthropic server; Team/Enterprise does not train according to commercial data commitments, and personal plan pricing page labels can exit training. Google Connector raw data is not trained, but copy/paste and responses containing its information can be affected by personal training settings.
Limitations: Each retention period of personal data has not been verified this time; organization-customized retention cannot be presumed to be personal default.
External measurements & hands-on results
Different methods, samples and dates. Not combined into one success rate.
External record ↗
- Evidence type
- single_user_bug_report_with_logs
Analyst notes
Key differences
- Unified dialogue, project memory and editable product collaboration; it has changed from local-based to mandatory cloud for Pro/Max new tasks.
- The cloud background and local operations are two layers: what can be continued after disconnection/shutdown is cloud work, not local device access.
- Old Cowork, unified Claude, @Claude, and Claude Code must be recorded in different modes.
Known unknowns
- Account-level actual migration coverage
- Is the complete agent product API open?
- Hard upper limit for personal task concurrency
- Personal independent email/mobile phone number/wallet
- Closed-loop coverage of consumption reservation payment
- Connection service any event trigger range
- Reproducible and comparable current product-level execution success rates
Source conflicts and how they were resolved
- topic: Scheduled running position; note: The new version of the cross-platform document clarifies Pro/Max new tasks in the cloud, but the scheduled manual setting section retains "requires local files or apps to only run locally"; it is layered by date/package/old tasks and is not abstracted into a global switch.; source ids: C05; C09; C10; source urls: https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile; https://support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-claude-cowork; https://support.claude.com/en/articles/14479288-claude-cowork-architecture-overview
- topic: Product mobile version and beta status; note: The cross-platform help page form still limits the mobile availability of some artifacts; the announcement on October 8th clarified that Docs/Slides/Design mobile editing will exit beta, adopt newer specific announcements, and retain the organization switch.; source ids: C05; C02; source urls: https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile; https://claude.com/resources/articles/dashboards-and-motion
- topic: Voice data time difference; note: The search cache still says that Cowork does not support voice; the real-time help page clearly unifies the Claude task dialogue to support voice, and the old Cowork/Code still only supports dictation.; source ids: C15; C01; source urls: https://support.claude.com/en/articles/11101966-use-voice-mode; https://claude.com/resources/articles/cowork-is-now-claude
Similar agents
All Claude (including Cowork execution) alternatives →