XDA put ChatGPT Work up against Claude Cowork last month and the surprise was that neither won on raw capability. The result came down to fit: how each agent slotted into an existing browser, inbox, and calendar without breaking flow. That is the real test for the best apps for AI work agents right now, since every tool on our shortlist can technically draft an email or click through a form. What separates them is permission scope, per-seat billing, memory across sessions, and how much they hallucinate when the browser reaches a login wall. We tested seven agents across Windows, macOS, and Linux for two weeks. The picks below reflect what stuck after that fortnight, not what looked good in a demo.
What to look for in an AI work agent
Action surface is the first cut. A chat window that summarises your inbox is not an agent. A tool that can open Gmail, read a thread, draft a reply, and pause for approval is. Ask which apps it touches out of the box, and whether it uses a native browser, a headless browser, or a Chromium extension you already trust.
Permission model matters more than model quality. Look for granular scopes (read-only, draft-only, send with confirmation) and a clear audit log. Anything that asks for full account access on day one is a policy problem waiting to happen.
Per-seat pricing rarely tells the full story. Agent runs burn ten to a hundred times more tokens than a chat turn, so the useful metric is what a working day of runs costs on your plan, not the sticker price.
Memory across sessions is what turns a demo into a workflow. Agents that forget your reporting structure or your CRM field names every morning will bleed hours to re-briefing. Same for the audit log: without one, you cannot trust what the agent did while you were in a meeting.
Quick comparison
| App | Best for | Deployment | Free tier | Starting price/mo | Access surface |
|---|---|---|---|---|---|
| Claude Cowork | Cross-app agent work | Web, Windows, macOS, Linux (via web) | Limited chat, no Cowork | $20 Pro | Anthropic sandbox browser |
| ChatGPT Agent | OpenAI-loyal teams | Web, Windows, macOS native | Limited chat | $20 Plus | Sandbox VM with browser |
| Microsoft Copilot | Microsoft 365 shops | Windows native, web | Copilot free tier | $30 M365 Copilot add-on | Windows, Office, Edge |
| Perplexity Comet | Browser-first research | Windows, macOS, Linux native | Full agent free | $5 Comet Plus | Native browser |
| Zapier Agents | SaaS-to-SaaS workflows | Web | 400 activities/mo | $33 Pro (annual) | 8,000+ app connectors |
| Notion AI | Teams that live in Notion | Web, Windows, macOS, Linux | Trial only | $20 Business | Notion databases, docs |
| Google Gemini for Workspace | Google Workspace teams | Web, Chrome extension | Starter side-panel | $14 Standard | Gmail, Docs, Sheets, Meet |
1. Claude Cowork, best for cross-app agent work
Anthropic’s Cowork spins up a sandbox browser on Anthropic servers, opens the sites you named, and works through the task with a live stream of what it is doing. In our runs it filed expense reports across Concur, pulled Q2 sales from HubSpot into a Google Sheet, and drafted follow-ups without asking for extra credentials mid-run. It went generally available on April 9, 2026, and Anthropic added cloud remote sessions in July so tasks keep running after you close the laptop.
Where it falls short: login walls still break it. Any site with two-factor or a captcha needs you to step in, and a Cowork run burns roughly 50 to 100 times the tokens of a chat turn, so Pro usage caps arrive quickly.
Pricing: Free chat plan has no Cowork access. Cowork is included on Pro at $20 per month, Max 5x at $100, Max 20x at $200, and every Team seat. Max 5x is the realistic floor for daily use.
Platforms: Web, plus Windows, macOS, and iOS clients. Linux users get it through the web app.
Download: claude.com/download
Bottom line: Get Claude Cowork if your work spans five browser tabs and a couple of SaaS tools, and if you can budget for Max 5x.
2. Perplexity Comet, best free browser agent
Comet is a full Chromium browser with an agent stitched into every tab. It reads the page you are on, opens new tabs when it needs to, and files results back into your notes or a Google Doc. Perplexity dropped the paywall on March 18, 2026, so the base agent, page summarisation, and Deep Research all run on the free tier across Windows, macOS, and Linux.
Where it falls short: it is a browser, so it inherits every quirk of browser-based agents (site rate limits, captcha loops, inconsistent behaviour on JavaScript-heavy dashboards). The agent also does not cross into your email or file storage without extra permissions.
Pricing: Free covers the full agent. Comet Plus is a $5 per month add-on for premium publisher content, included for Pro and Max subscribers. Perplexity Pro at $20 unlocks longer research runs.
Platforms: Native builds for Windows, macOS, Linux, iOS, and Android.
Download: perplexity.ai/comet
Bottom line: Try Comet first if you want an agent that costs nothing and lives where you already browse.
3. ChatGPT Agent, best for OpenAI-loyal teams
Agent mode lives inside ChatGPT and runs in a virtual machine with its own browser, terminal, and code interpreter. In our tests it built a competitor pricing sheet from ten vendor sites, wrote a Python script to clean it, and dropped the result into a shared Drive folder. Business seat pricing dropped to $20 per seat per month on annual billing in April 2026, matching Plus.
Where it falls short: since July 6, 2026, agent runs draw down workspace credits on top of the seat price. A typical run costs 5 to 25 credits, which adds a variable line to the invoice that seat pricing alone does not capture.
Pricing: Plus at $20 per month includes Agent mode. Business at $20 per seat per month on annual billing (monthly billing runs $25 to $30 with a two-seat minimum). Enterprise starts around $60 per seat with a 150-seat floor.
Platforms: Web, native Windows, native macOS. Linux is web-only.
Download: openai.com/chatgpt/download
Bottom line: Pick ChatGPT Agent if your team already pays for Plus or Business and you want the deepest tool integrations in a single subscription.
4. Microsoft Copilot, best for Microsoft 365 shops
Copilot on Windows 11 sits in the taskbar and reaches into Outlook, Word, Excel, Teams, and Edge. Agent Studio (rebranded from Copilot Studio) lets ops teams build agents that read shared mailboxes, run against SharePoint, and post into Teams channels. Microsoft added Copilot Cowork on June 16, 2026, which bills per interaction and covers the long-running browser agent use case.
Where it falls short: the pricing is three meters stacked. You pay the M365 base seat, then $30 per seat for Copilot, then Copilot Studio capacity packs at $200 per month for 25,000 credits, plus pay-as-you-go at $0.01 per credit for anything else. Budgets get complicated fast.
Pricing: M365 Copilot at $30 per seat per month on annual commit, on top of a qualifying M365 license. Base-suite prices rose July 1, 2026, so the all-in floor is around $69 per seat on E3 or $90 on E5.
Platforms: Windows 11 native, web, and inside every Microsoft 365 app.
Download: microsoft.com/copilot
Bottom line: Buy M365 Copilot only if you already run Microsoft 365 at scale and can absorb the metered credit bill on top of seats.
5. Google Gemini for Workspace, best for Google Workspace teams
Gemini shows up as a side panel in Gmail, Docs, Sheets, Slides, and Meet, and the standalone Gemini Enterprise product adds a no-code agent builder with pre-built agents for Deep Research, NotebookLM, and Data Insights. In our tests it summarised a 40-message thread, drafted a reply, and pulled numbers from three Sheets into a briefing doc without leaving the Workspace surface.
Where it falls short: the agentic behaviour outside Workspace apps is thinner than Claude Cowork or ChatGPT Agent. If your workflow leans on Salesforce, ServiceNow, or Jira, you probably want Gemini Enterprise rather than the bundled version.
Pricing: Bundled with Workspace at roughly $7 Starter, $14 Standard, $22 Plus per user per month on annual billing. Gemini Enterprise (standalone agent platform) starts at $21 per seat per month Business, $30-plus Standard, plus consumption charges.
Platforms: Web across every Workspace app, plus a Chrome extension. No native desktop client.
Download: workspace.google.com/products/gemini
Bottom line: Choose Gemini for Workspace if your team already runs on Gmail and Docs and you want AI without adding a second billing relationship.
6. Zapier Agents, best for SaaS-to-SaaS workflows
Zapier Agents sit on top of the 8,000-plus app connectors Zapier already has, so an agent can watch a Slack channel, file a Jira ticket, tag it in Linear, and email the summary to a stakeholder without you writing a single zap. It is the fastest way to wire an agent into tools that do not have their own AI offering.
Where it falls short: it does not drive a browser. Anything that requires clicking through a web app without an API is out of scope, which rules out a chunk of the tasks Claude Cowork handles natively.
Pricing: Agents bill separately from core Zaps, metered by activities. Free tier gets 400 activities per month. Pro is $33.33 per month billed annually for 1,500 activities. Higher tiers scale by activity volume.
Platforms: Web only. Runs in any browser on Windows, macOS, and Linux.
Download: zapier.com/agents
Bottom line: Reach for Zapier Agents when the task is stitching together APIs, not driving a browser.
7. Notion AI, best if your work lives in Notion
Notion Custom Agents shipped on May 4, 2026, and they run against your workspace databases, docs, and connected sources (Slack, Google Drive, GitHub). We wired one to a weekly ops review database that reads the last week of updates, drafts the summary block, and links every claim back to the source page.
Where it falls short: agents only reach outside Notion through its Connections list, so if your critical data is in a system Notion has not added a connector for, you are stuck. Credits do not roll over, which punishes light users at the tier boundary.
Pricing: Full Notion AI is bundled with Business at $20 per user per month ($24 monthly). Custom Agents draw down workspace-pooled credits priced at $10 per 1,000 credits, resetting monthly.
Platforms: Web, native Windows, native macOS, native Linux client.
Download: notion.so/desktop
Bottom line: Turn on Notion AI Custom Agents if the team already documents everything in Notion, and skip it otherwise.
How to pick the right one
If you want the strongest cross-app browser agent and can afford a real subscription, get Claude Cowork on Max 5x. It handles the widest range of tasks with the fewest permission surprises.
If your budget is zero, install Perplexity Comet and use it as your default browser for a week. The free tier covers the full agent, and you can add Comet Plus for $5 later if you want premium publisher content.
If your company runs on Microsoft 365, Copilot is the least friction path even with the metered credit bill on top of seats. Same logic for Google Workspace shops and Gemini for Workspace, only cheaper.
If your work is API-heavy (CRMs, ticketing, project management) and light on browser clicks, Zapier Agents will beat every browser-native agent on reliability. If most of your operating docs already live in Notion, Notion Custom Agents will feel more useful than a general-purpose agent. If you were going to run a lot of deep research reports for legal, consulting, or finance work, look at You.com ARI on the Max plan or the Research API.
FAQ
What is an AI work agent?
An AI work agent goes beyond a chat window. It takes a goal, opens the apps and websites needed to reach that goal, performs actions (clicks, form fills, file writes), and reports back with what it did. The best ones pause for approval before anything destructive and keep an audit log of every step.
How is ChatGPT Agent different from Claude Cowork?
Both spin up a sandbox environment and drive a browser to complete tasks. In XDA’s head-to-head the differences came down to fit rather than raw capability. ChatGPT Agent leans on a wider set of built-in tools (code interpreter, terminal), while Claude Cowork is stronger at long multi-step browser tasks and now runs remote sessions that survive after you close your laptop.
Are AI work agents safe to give access to my email?
Only with granular permissions and a visible audit log. Every agent on this list supports scoped access (read-only, draft-only, send with approval). We would not run any of them against a personal Gmail with a saved payment method attached, but a work inbox with SSO and 2FA is a reasonable starting point.
What does an AI work agent actually cost per month?
Sticker prices start at $0 (Perplexity Comet) and $20 (Claude Pro, ChatGPT Plus, Notion Business), but agent runs burn tokens or credits fast. Plan on Max 5x ($100 per month) for daily Claude Cowork use, Business plans with capacity packs for Microsoft Copilot, and $33 per month plus activity overage for Zapier Agents.
Do any AI work agents run locally on my machine?
Not the ones tested here. Every agent on this list runs its browsing session in a cloud sandbox, either the vendor’s or a workspace-scoped VM. Local agent frameworks exist (Auto-GPT forks, LangChain, LM Studio agent workflows), but they lack the polish and permission model of the hosted tools.
Which AI work agent has the best free tier?
Perplexity Comet, comfortably. It ships as a full browser with the agent, page summarisation, voice mode, and Deep Research all free after March 18, 2026. The only paid add-on at the individual tier is Comet Plus at $5 per month, and Pro and Max subscribers get that included.
