Two always-on AI agents are now competing to do your work while you are away from the screen. OpenAI’s Dots lives inside the tools your team already uses, keeps watching them, and messages you in Slack or Teams when something needs attention. Anthropic’s Claude Cowork starts from your own files instead: point it at a folder, describe the outcome, and it produces finished documents, spreadsheets and presentations. They are built for different jobs, and picking the wrong one wastes money or hands over access you did not mean to grant. This head-to-head compares Dots and Cowork task by task, grounded in company launch material and dated independent reporting checked 22 October 2026. It is desk-validated, not hands-on tested — and that matters here, because no controlled task-by-task benchmark of the two agents exists anywhere we could find.
What is each agent, in one paragraph?
Dots is OpenAI’s always-on personal agent for ChatGPT users. Launched at DevDay on 29 September 2026, each Dot gets its own OpenAI-managed cloud computer with a virtual browser, so it keeps working while your devices are off. You set a goal and a boundary, and it follows up as things change; a proactive mode does read-only research when idle — enforced in code, so it cannot send messages or change anything in that mode — while anything that acts goes through Custom Rules and an auto-review checker. Dots lives in ChatGPT’s desktop, web and mobile apps, in Slack and Microsoft Teams, and reaches more than 4,000 apps through plugins. It is sold only on higher paid plans, starting at ChatGPT Pro at $100 a month.
Cowork is Anthropic’s file-based agent for knowledge work. After a research preview in January 2026 and general availability on 9 April 2026, it was merged into main Claude on 16 September 2026 (“Cowork is now just Claude”), and since 6 October 2026 new tasks run by default in an isolated cloud environment on Anthropic’s servers. The model is simple: point it at selected folders on your computer, describe the task in plain language, and it plans and executes multi-step work through an orchestrator with parallel sub-agents, returning polished files — Excel spreadsheets with real formulas, PowerPoint decks, Word documents, PDFs. There is no separate Cowork subscription: it is included in paid Claude plans from Pro at $20 a month. It asks before acting by default, and approval modes run from Manual to Auto to Skip-all.
Cowork vs Dots at a glance
The table separates what is shipped today from what is merely announced. Prices are dated and attributed; treat vendor claims as claims, not tested facts.
| OpenAI Dots | Claude Cowork | |
|---|---|---|
| Best for | Standing work that watches connected apps and reaches you where you work | Turning your own files into finished documents, decks and spreadsheets |
| Launched | 29 September 2026 (DevDay) | Preview 12 January 2026; GA 9 April 2026; merged into Claude 16 September 2026 |
| Model | GPT-6 Astra only | Claude 5.5 family (Opus 5.5, Sonnet 5.5, Haiku 5.5) plus Fable 5.1, Mythos 5.1, Opus 5; effort levels Low to Max |
| Cheapest way in | ChatGPT Pro 100, $100/month; Business Premium $125/seat/month | Claude Pro, $20/month — a fifth of Dots’ entry price |
| Where you use it | ChatGPT desktop, web and mobile; Slack; Microsoft Teams | Desktop, web, mobile, Claude-in-Chrome side panel |
| Isolation | Own OpenAI-managed cloud computer per Dot | Isolated cloud environment; sandboxed to folders you allow |
| Who sets the boundary | You via Custom Rules; workspace admin can set the outer boundary (enterprise beta off by default) | You, by folder and tool selection; asks before acting by default |
| Regions | Pro excludes the EEA, Switzerland and the UK; Business Premium across supported ChatGPT regions | Plan-based rollout (Pro/Max first); no full country list published |
A note on Anthropic’s 21 October 2026 news: the launch-week “Claude Duet” is a separate always-on proactive assistant (Opus 5.5, waitlisted research preview, US Pro/Max only), not Cowork. It does not change anything in this comparison.
What does each agent actually cost?
Dots costs more because it is sold as a seat upgrade. OpenAI’s October 2026 pricing puts entry at Pro 100 ($100/month), Pro 200 ($200/month) and Pro 500 ($500/month), with Business Premium at $125 per seat per month ($100 on annual billing, two-seat minimum). The first Dot is included per eligible user. Free, Go, Plus and Business Standard plans are excluded from the rollout. Two figures were not published at launch and are still missing: what Dot usage costs after the launch-month grace period (Dot conversations did not count toward plan limits at launch), and what a second Dot costs — teams-of-dots is planned, not shipped.
Cowork takes the opposite approach: there is no Cowork subscription at all. It rides on your Claude plan: Pro at $20/month ($17 on annual billing), Max 5x at $100/month, Max 20x at $200/month, Team Standard at $25/seat/month, Team Premium at $125/seat/month, and Enterprise from $20/seat/month plus API usage. The tiers buy usage capacity, not different features. There is no free-tier Cowork.
The useful test is cost per completed task, not the rate card. Both agents are meant to work continuously, and Anthropic’s own pricing notes warn that agentic multi-step tasks burn through usage allowances faster than chat. Before committing a team, run the same three real tasks through your shortlist and record how many finished correctly per dollar — including the human review time each one still needed.
Six everyday tasks, scored side by side
No controlled benchmark exists — the table below scores each task on documented capability, not on hands-on testing. Every independent comparison we found (layer3labs, tooljunction) is spec- and architecture-based; layer3labs states explicitly that real-world Dots reports were “not in yet”. Where a task has no verified example for one agent, the table says so.
| Everyday task | OpenAI Dots | Claude Cowork | Honest read |
|---|---|---|---|
| Inbox triage | Documented: proactive read-only research can watch connected email; sending requires approval via Custom Rules. A Dot can connect your personal email account (no standalone Dot email yet) | Asks-before-acting by default; Gmail via 6,000+ Claude Marketplace connectors — but no verified inbox-triage workflow example found | Dots has the documented always-on watching pattern; Cowork’s email flow is connector-dependent and untested in the record |
| Meeting prep | Documented example: interview transcripts turned into show notes plus draft social posts | M365 and Google Workspace connectors cover the raw materials; no verified meeting-prep example found | Dots has the only documented example in this category; otherwise unconfirmed |
| Research synthesis | Read-only proactive research plus its own browser; OpenAI’s example roles include research-heavy work like a scientist rerunning analyses | Verified strength: “papers plus notes become a literature review with source organisation and citation”, with parallel sub-agents | Both verified; Cowork’s synthesis framing is the most concretely documented |
| Calendar scheduling | Reminders and recurring checks; Teams integration — but no verified calendar-booking example found | No verified calendar example found (the scheduling claim belongs to Duet, a separate product) | Unconfirmed for both — neither agent should be bought for booking meetings |
| Finished documents and spreadsheets | Delegates: “delegating complex document production to connected apps or ChatGPT Work” | Verified core strength: native Excel .xlsx with formulas, PowerPoint, Word, PDF; Claude Docs and Slides beta exporting to PPT/PDF; “local CSV/Excel to formatted reports” | Cowork wins on verified facts — native deliverables are its documented centre of gravity |
| Team-delegated multi-step projects | Parallel jobs, persistent identity and memory, Slack/Teams delivery; teams-of-dots planned, not shipped; specialist Dots in enterprise pilots | Verified: orchestrator plus parallel sub-agents, scheduled recurring tasks, Cowork Projects with own instructions and memory, Team/Enterprise RBAC, SCIM, spend limits and usage analytics | Both verified; Cowork has the more concrete team machinery in-market today; Dots has the ambient, always-on channel presence |
Pick Cowork if…
- Your job starts with “here are my files, turn them into a finished report” — file-grounded knowledge work is Cowork’s documented home ground.
- The deliverable is a spreadsheet with real formulas, a deck or a formatted document, not a message.
- You want the cheapest serious entry point: $20/month, five times cheaper than Dots’ cheapest tier.
- You are comfortable with the known limits: audit logs do not currently capture Cowork activity, a cloud session reaches your local folders only while Claude Desktop is open, and local tasks do not use memory.
Pick Dots if…
- Your job is ongoing, not one deliverable: watching connected apps, recurring checks, follow-ups that arrive in Slack or Teams.
- A workspace admin must set and audit the outer boundary before anyone delegates work — Dots’ enterprise beta is off by default until an admin enables it.
- Your team already pays for ChatGPT Pro or Business Premium and works inside Slack or Microsoft Teams.
- You accept the unknowns: ongoing usage pricing after the launch month is unpublished, a second Dot has no published price, and Dots is the younger product — Cowork has been generally available since April 2026.
How this fits the three-way picture
This article stays strictly Dots-vs-Cowork. If Meta’s Muse is also on your shortlist — the free consumer agent for personal errands — read our three-way comparison, Muse vs Dots vs Cowork: Which Agent Fits You?, which adds Muse’s reach-and-free-tier angle to the same decision. Start every agent read-only, widen access only after it has earned trust on low-stakes tasks, and never give send, publish or spend rights without a human approval step.
FAQ
Do I need a paid plan for Dots?
Yes. Dots launched on paid plans only, starting at ChatGPT Pro at $100/month or Business Premium at $125 per seat. Free, Go, Plus and Business Standard plans are excluded. Each eligible user starts with one Dot; pricing for additional Dots has not been published.
Is Cowork a separate subscription?
No. Since 9 April 2026, Cowork is included in paid Claude plans, starting at Pro $20/month ($17/month on annual billing), with Max at $100 and $200/month buying more usage capacity rather than different features. There is no free-tier Cowork.
Which agent is safer for company data?
On documented controls alone, Dots, because a workspace admin can set the outer boundary and enterprise use is off by default. Cowork is next: it is scoped to folders you choose and asks before acting by default. One verified Cowork limitation cuts the other way — audit logs do not currently capture Cowork activity. Neither agent should be given send, publish or spend rights without a human approval step.
Can either agent work while my computer is off?
Dots can: each Dot runs on its own OpenAI-managed cloud computer and keeps working with your devices off. Cowork is more qualified: since 6 October 2026, new Pro/Max tasks run in the cloud by default, and scheduled tasks run without your computer — but a cloud session can only reach your local folders while Claude Desktop is open.
Is Claude Duet the same as Cowork?
No. Claude Duet, launched 21 October 2026, is a separate always-on proactive assistant built on Opus 5.5, currently a waitlisted research preview for US Pro/Max subscribers. Cowork remains the file-based work capability inside Claude; Duet does not change this comparison.
Sources and methodology
This comparison rests on company launch material and dated independent reporting, checked 22 October 2026. Dots facts (29 September 2026 DevDay launch, GPT-6 Astra, cloud computer, proactive read-only research, Slack/Teams presence, Pro 100/200/500 and Business Premium pricing, plan exclusions, EEA/Switzerland/UK exclusion) come from launch-week reporting summarised by layer3labs.io and apex36tech.com in September–October 2026; ongoing usage terms and extra-Dot pricing are flagged as unpublished where the sources say so. Cowork facts (January preview, 9 April 2026 GA, 16 September 2026 Unified Claude merge, 6 October 2026 cloud-by-default change, folder-scoped execution, M365 and Google Workspace connectors, native .xlsx/PowerPoint/Word/PDF outputs, Pro/Max/Team/Enterprise pricing, audit-log limitation) come from Anthropic-adjacent reference material at cowork.bruniaux.com and futureability.com, current to October 2026. Anthropic’s analysis of 1.2 million anonymised sessions across 600,000+ organisations (33.4% business processes and operations, 16.4% content creation, 8.7% software development) is company analysis, reported by Tech Times in July 2026. The layer3labs.io Dots-vs-Cowork comparison (September 2026) states explicitly that its verdicts are spec-based, with real-world Dots reports not yet in; no controlled task-by-task benchmark exists, so the six-task table scores documented capability only — it is not a test result. Where sources conflicted, the more conservative, company-sourced claim was used and the conflict noted. Read the primary comparisons here: layer3labs.io’s Dots vs Claude Cowork, cowork.bruniaux.com’s Cowork guide, and apex36tech on Dots pricing and safety. Internal reading: our three-way Muse vs Dots vs Cowork comparison and our testing methodology. This article was desk-validated, not hands-on tested.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.