ChatGPT Work vs Claude: which is better for 12 assistant jobs?
Short answer: ChatGPT Work suits work that starts on its own, because OpenAI documents event triggers such as a new Gmail message. Claude documents more on Gmail actions and privacy. Price is the same at the entry tier: $20 a month. If you would rather have an assistant you cc on email threads and text, rather than a chat agent, see OpenAssistant.
Both products in this comparison have been renamed, so first, what we are actually comparing.
ChatGPT agent is gone. OpenAI’s help page for it now opens with “ChatGPT agent is no longer available” and points people to ChatGPT Work for multi-step jobs and to the cloud browser for web tasks (OpenAI help). OpenAI describes Work as “an agent designed for longer, multi-step work and finished deliverables” (OpenAI help). So the OpenAI side of this article is ChatGPT Work with the cloud browser.
Claude Cowork is now just Claude. Anthropic’s Cowork page carries the banner “Claude Cowork is now just Claude” (Claude). Anthropic announced the merge of Cowork and chat on 16 September 2026, with Pro and Max first and Team and Free to follow (Anthropic). So the Anthropic side is Claude with Cowork features, wherever the merge has reached a given plan.
Both products do the same kind of job: you give a goal, and the agent works across your files, apps and the web until it has a result. This article lines up what each vendor publicly documents for 12 of the 25 jobs in our evaluation framework.
The 12 jobs
| # | Framework job | What to look for in the documentation |
|---|---|---|
| 1 | Job 5: Triage a busy inbox | Read and write access to a real mailbox through a connector |
| 2 | Job 6: Draft replies in the user’s voice | Drafting, and whether sending needs approval |
| 3 | Job 7: Find the thread you forgot to answer | Reading whole threads, not single messages |
| 4 | Job 8: Send a daily inbox digest at a set time | Scheduled work that runs while you sleep |
| 5 | Job 11: Write a prep brief for an outside meeting | Calendar plus email plus web |
| 6 | Job 12: Send a daily agenda | Scheduled work with fresh data |
| 7 | Job 16: Turn a screenshot into a calendar event | Images, time zones, calendar writes |
| 8 | Job 17: Research a company and write a brief | Web access and sourcing |
| 9 | Job 18: Build a spreadsheet from messy notes | Producing a real file |
| 10 | Job 19: Plan a trip and stop before paying | Browser work and stopping at the right step |
| 11 | Job 22: Audit what it can see, then revoke it | Permissions and admin controls |
| 12 | Job 23: Respect “ask me first” | Approval controls |
We left out the framework’s phone, texting and own-email jobs. Neither product documents its own phone number or its own email address on the pages listed here, so those jobs would only show a missing feature.
What each product says it can do
All facts below come from the vendor’s own pages. This is a fact table, not a ranking.
| ChatGPT Work | Claude (Cowork) | |
|---|---|---|
| Entry plan | Plus, $20 a month (pricing) | Pro, $20 billed monthly or $17 a month billed annually (pricing) |
| Business seat | Business Standard, $25 monthly or $20 annually per seat, 2 to 200 users (pricing) | Team Standard, $25 monthly or $20 annually per seat, 2 to 150 users (pricing) |
| Higher tiers | Pro at $100, $200 or $500 a month, monthly billing only (OpenAI help) | Max at $100 (5x) or $200 (20x) a month (Cowork page) |
| Where it runs | All plans on desktop; Plus and up on web and mobile (Work page) | Desktop on all paid plans; web and mobile on Pro, Max and Team (Claude help); the product page calls web and mobile beta (Cowork page) |
| Free plan | “Limited” Work access on desktop (pricing) | Cowork is paid plans only (Claude help) |
| Its own browser | Cloud browser on a separate cloud computer, paid plans except Go (OpenAI help) | Built-in browser in the desktop app, rolling out to Pro, Max and Team (Anthropic); Claude in Chrome on all paid plans (Claude in Chrome) |
| Keeps working when your laptop is closed | Yes, per OpenAI (OpenAI help) | Yes, per Anthropic (Cowork page) |
| Scheduled work | One-time or recurring tasks; 5 active on Plus, 10 on Business, 15 on Pro and Enterprise (OpenAI help) | Hourly, daily, weekly or weekday tasks on all paid plans; runs while your computer sleeps unless it needs local files (Claude help) |
| Triggered by events | Event triggers such as a new Gmail message, on Plus and up (OpenAI help) | Not stated on the pages linked here |
| Gmail and Google Calendar | Connects Gmail, Calendar and Drive as Google apps (OpenAI help) | Gmail send, reply and forward, with approval before each by default; Calendar create, update, delete and find shared free time (Claude help) |
| Outlook | Outlook Email and Calendar apps; some actions need Microsoft admin consent and a workspace admin to enable them (OpenAI help) | Microsoft 365 connector on all plans for work accounts; write tools only if an admin enables them; personal Outlook.com accounts cannot connect (Claude help) |
| Approval before acting | Asks for confirmation before consequential steps like a booking or payment (OpenAI help); Plan mode shows a plan for approval first (Work page) | Manual approval by default, plus Auto and Skip modes; always asks before permanently deleting files (Claude help, safety page) |
| Training on your data, consumer tier | May train on Free, Plus and Pro content unless you opt out (OpenAI help) | Your choice since the 28 August 2025 terms update; 5 year retention with training on, 30 days with it off (Anthropic) |
| Training on your data, business tier | Not by default on Business and Enterprise (OpenAI help) | Not by default on commercial plans (Anthropic privacy) |
| Certifications listed | Not stated in the sources linked here | SOC 2 Type II, ISO 27001 and ISO 42001 (Anthropic trust center) |
Two details stand out.
First, Anthropic tells you not to schedule the risky stuff. Its safety page says “Don’t schedule tasks that access sensitive files, send messages on your behalf” and similar (Claude help). That bears on the scheduled jobs below (digests and agendas): a schedule that only delivers a summary to you avoids the sending part of that advice.
Second, usage is metered differently. OpenAI says Work “follows the same usage structure as Codex” rather than a fixed task count (OpenAI help). Anthropic says Cowork “consumes limits faster than Chat” (Cowork page). Neither vendor publishes a number of agent tasks per month on the pages linked here, so you cannot work out from public information how many of these jobs a plan will cover.
What each vendor documents, job by job
Each job below says what the job needs, then what the two vendors publicly document. Where a vendor’s pages are silent, we write “Not stated”.
1. Triage a busy inbox (Job 5)
Archiving and labeling need write access to the mailbox. Claude’s Gmail connector page says it can “Manage email organization with labels and threads” (Anthropic). OpenAI says its Gmail app can “create reply drafts, and archive, label, or move messages to Trash” when asked (OpenAI). Both document mailbox clean-up.
2. Draft replies in the user’s voice (Job 6)
OpenAI documents reply drafts (OpenAI). Claude’s Gmail connector lists send, reply and forward, with approval before each by default (Anthropic). The sources linked here do not say whether Claude saves drafts in the mail client, and none of the OpenAI pages listed here lists sending Gmail. The sources used for this article do not mention learning your tone from past mail for either product.
3. Find the thread you forgot to answer (Job 7)
Both products connect to mail, which is the precondition. The sources used for this article do not describe a feature for spotting unanswered threads. Not stated for either product.
4. Send a daily inbox digest at a set time (Job 8)
Both document scheduling. OpenAI: one-time or recurring tasks, 5 active on Plus, 10 on Business, 15 on Pro and Enterprise (OpenAI help). Anthropic: hourly, daily, weekly or weekday tasks on all paid plans, running while your computer sleeps unless the task needs local files (Claude help). How a scheduled result is delivered (email or message) is not stated in the sources used here.
5. Write a prep brief for an outside meeting (Job 11)
This job needs calendar, email and web together. Both connect Gmail and Calendar (see the table), and both have a browser (OpenAI’s cloud browser; Anthropic’s built-in browser and Claude in Chrome). The sources used here do not name a prep-brief feature for either product.
6. Send a daily agenda (Job 12)
Same scheduling facts as Job 4. The extra point is freshness: an agenda is only useful if it reflects the calendar as it stands. The sources used here do not say how fresh scheduled results are for calendar data.
7. Turn a screenshot into a calendar event (Job 16)
The calendar side is documented. Claude lists Calendar create, update and delete (Anthropic), and OpenAI connects Calendar as a Google app (OpenAI help). The sources used here do not document reading a screenshot and converting its time zone into a calendar event.
8. Research a company and write a brief (Job 17)
Both have web access through their own browsers (see the table). OpenAI describes Work as built for “finished deliverables” (OpenAI help). The sources used here make no claim about citing sources or about saying “not found” when a fact is missing.
9. Build a spreadsheet from messy notes (Job 18)
OpenAI’s “finished deliverables” wording points at file output. The sources linked here do not name spreadsheets for either product, or saving to Drive or OneDrive. Not stated by either.
10. Plan a trip and stop before paying (Job 19)
Both vendors say their agents stop before payments. OpenAI’s cloud browser asks for confirmation before steps like confirming a booking or paying (OpenAI help), and Anthropic says purchases “wait for you” in Claude in Chrome (Claude in Chrome). These are the vendors’ own statements.
11. Audit what it can see, then revoke it (Job 22)
Both connect to Google and Microsoft accounts, and both put limits on Microsoft access: OpenAI says some Outlook actions need Microsoft admin consent (OpenAI help), and Anthropic says Microsoft 365 write tools need an admin to turn them on (Claude help). Anthropic lists SOC 2 Type II, ISO 27001 and ISO 42001 (trust center). Steps for revoking access are not described on the pages linked here, and we did not research OpenAI’s certifications.
12. Respect “ask me first” (Job 23)
This is where the two documentation sets are most concrete. OpenAI: confirmation before consequential steps, and Plan mode shows a plan for approval first. Anthropic: manual approval by default, Auto and Skip modes, approval before each Gmail send, and a prompt before permanently deleting files. Both vendors describe approval steps as part of normal use.
Who documents what
The criteria follow our evaluation framework. This is an editorial judgment from the linked facts above, not a score.
| Criterion | Which vendor documents more | Reasoning |
|---|---|---|
| Documented capabilities cover the job | Close. ChatGPT Work for event-driven work; Claude for Gmail actions | OpenAI documents event triggers and Claude does not. Claude documents Gmail send, reply and forward with approval, and OpenAI’s linked pages list drafts, archive, labels and trash. |
| What the vendor says about quality | Neither | Neither vendor publishes results for these jobs on the pages linked here. |
| Price for what you get | Equal | Both are $20 a month at the entry paid tier and $25 monthly or $20 annually on the business seat. Claude’s annual Pro price is $17 a month. |
| Where it works | Close | Both cover Gmail, Google Calendar and Microsoft 365, with admin gates on Microsoft. OpenAI runs on web and mobile from Plus; Claude’s web and mobile are labeled beta on its product page. |
| Privacy, security and admin | Claude, on what is published | Anthropic lists three certifications and asks consumer users to choose on training. OpenAI may train on consumer content unless you opt out. On business seats neither trains by default. |
What the documentation suggests
- ChatGPT Work looks better set up for work that starts on its own. It documents event triggers, such as a new Gmail message, on Plus and up. Claude’s pages describe only time-based schedules.
- They document different Gmail powers. Claude’s Gmail connector lists send, reply and forward, with approval before each by default, plus labels (Anthropic). OpenAI’s Gmail app lists drafts, archive, labels and trash (OpenAI), but none of the OpenAI pages linked here lists sending Gmail.
- For Outlook users, both need an admin. OpenAI says some Outlook actions need Microsoft admin consent, and Anthropic says Microsoft 365 write tools need an admin to turn them on. If you are on a work Microsoft account and not the admin, check with IT before you count on either one to send or edit.
- Default privacy differs on the consumer tier. OpenAI may train on Plus content unless you opt out. Anthropic asks you to choose. On business seats, neither trains on your data by default.
Related reading
- Gemini vs Microsoft Copilot for work: which fits your team?: Gemini and Microsoft 365 Copilot compared on 12 work jobs.
- Why we recommend OpenAssistant, with 11 alternatives compared: twelve assistants ranked on documented jobs, price, channels and privacy.
- How to Use ChatGPT Agent for Errands (Now ChatGPT Work): a step-by-step run through ChatGPT Work errands, with Claude in Chrome as the alternative.
- How to Use Claude Cowork for Admin Work: a step-by-step run through Claude Cowork tasks, files and scheduled tasks.
Frequently asked questions
Is ChatGPT Work or Claude better for work tasks?
It depends on the job. ChatGPT Work documents event triggers, such as a new Gmail message, on Plus and up. Claude documents Gmail send, reply and forward with approval, and lists three certifications. Both cost $20 a month at the entry tier.
How much do ChatGPT Work and Claude cost?
Both are $20 a month at the entry paid tier (ChatGPT Plus, Claude Pro, which is $17 a month billed annually). Business seats are $25 monthly or $20 annually on both.
Can they run tasks on a schedule?
Yes. OpenAI documents one-time or recurring tasks, 5 active on Plus. Anthropic documents hourly, daily, weekly or weekday tasks on all paid plans.
Do ChatGPT and Claude train on my data?
On consumer plans, OpenAI may train on content unless you opt out, and Anthropic asks you to choose. On business seats, neither trains on your data by default.
Want an assistant you can text?
ChatGPT Work and Claude are chat agents you open and prompt. OpenAssistant works by email, text and phone instead, and you can try it free with 1,000 credits at signup. See its pricing for Pro and Team.