Agentuna Agentuna

Best general AI agent for multi-step jobs end to end

Short answer: ChatGPT scores highest in our provisional ranking because ChatGPT Work (paid plans, launched July 2026) takes multi-step research and returns finished files, alongside chat, files and a strong free tier in one app (Plus $20/mo). Its old agent mode was retired in August 2026, and its cloud browser cannot log in to sites or pay. Manus is the purpose-built general agent: give it a goal and it researches, builds the document, slides or website in its own cloud computer (free credits, Standard $20/mo). Claude, with Cowork now inside the app on Pro from $20/mo, follows. Scores are editorial estimates, not test results.

One agent to plan and finish a multi-step job end to end.

Last checked: 2026-10-18 Score: provisional Demand signal: No X count yet; added for the agent focus, 18 Oct 2026
Scores are editorial estimates from public information. Hands-on tests have not started yet. Methodology

Scores are provisional editorial estimates: price and free-tier scores are read from each tool's official pricing page on the date shown; quality, ease and trust come from the vendors' own docs and product pages. No hands-on tests were run for this page.

The ranking

  1. #1

    ChatGPT

    ChatGPT Work: multi-step research and finished files on paid plans; its cloud browser cannot log in or pay. Plus $20/mo.

    Free tier From $8/mo Price from the official page · Not checked yet: Arabic support
    Score breakdown (provisional)
    Output quality (weight 30) 8/10 Arabic support (weight 5) 7/10 Price / value (weight 15) 7/10 Free tier (weight 10) 6/10 Ease of use (weight 20) 9/10 Reliability & trust (weight 20) 8/10
    78/100 · provisional
  2. #2

    Manus

    Purpose-built general agent with its own cloud computer; credits per task, Standard $20/mo.

    Free tier From $20/mo Price from the official page · Not checked yet: Arabic support
    Score breakdown (provisional)
    Output quality (weight 30) 9/10 Arabic support (weight 5) 5/10 Price / value (weight 15) 6/10 Free tier (weight 10) 6/10 Ease of use (weight 20) 8/10 Reliability & trust (weight 20) 6/10
    73/100 · provisional
  3. #3

    Claude

    Cowork lives in the Claude app on Pro, Max and Team; Pro from $20/mo.

    Free tier From $20/mo Price from the official page · Not checked yet: Arabic support
    Score breakdown (provisional)
    Output quality (weight 30) 8/10 Arabic support (weight 5) 7/10 Price / value (weight 15) 7/10 Free tier (weight 10) 4/10 Ease of use (weight 20) 7/10 Reliability & trust (weight 20) 8/10
    72/100 · provisional
  4. #4

    Genspark

    AI workspace with a Super Agent that researches and produces slides, documents and more.

    Free tier From $24.99/mo Price from the official page · Not checked yet: Arabic support
    Score breakdown (provisional)
    Output quality (weight 30) 8/10 Arabic support (weight 5) 5/10 Price / value (weight 15) 5/10 Free tier (weight 10) 6/10 Ease of use (weight 20) 8/10 Reliability & trust (weight 20) 6/10
    68/100 · provisional
  5. #5

    Microsoft 365 Copilot

    Copilot inside Excel, Word, PowerPoint and Outlook.

    No free tier From $9.99/mo Price from the official page · Not checked yet: Arabic support
    Score breakdown (provisional)
    Output quality (weight 30) 7/10 Arabic support (weight 5) 6/10 Price / value (weight 15) 5/10 Free tier (weight 10) 3/10 Ease of use (weight 20) 7/10 Reliability & trust (weight 20) 9/10
    67/100 · provisional
  6. #6

    Kortix

    Open-source AI operating system: agents with their own cloud computer, skills and company memory in one git repo.

    Free tier From $50/mo Price from the official page · Not checked yet: Arabic support
    Score breakdown (provisional)
    Output quality (weight 30) 7/10 Arabic support (weight 5) 4/10 Price / value (weight 15) 5/10 Free tier (weight 10) 6/10 Ease of use (weight 20) 5/10 Reliability & trust (weight 20) 5/10
    57/100 · provisional

Public accounts

What people and companies have publicly reported doing with these tools, in their own numbers, with the source. These are not our tests.

Nick Davidov (Bay Area investor) · Claude

Tidy a desktop with Claude Cowork (and what went wrong)

What they did
Nick Davidov asked Claude Cowork to organize his wife's desktop, including removing temporary Office files. Per t3n, it tried to merge folders named "Photos" and "photos" and, since macOS ignores case, deleted both with a terminal command.
How it went
About 15 years of family photos were wiped and were not in the Trash. Apple support walked him through an iCloud restore point that brought most of them back. The incident was on 7 February 2026 and the report is his, relayed by press.
Worth knowing
One user's account, relayed by outlets such as t3n and NDTV and listed as an alleged incident in the AI Incident Database. Give agents a copy of the files, keep a backup and read the plan before approving deletions.

t3n.de ↗ · · News report

Full example with a prompt to try

Simon Willison (newsletter) · Claude

Ask Claude Cowork which of your drafts are ready to publish

What they did
Simon Willison pointed Claude Cowork at his blog-drafts folder and asked it to find drafts from the last three months, check whether each was already published on his site, and suggest the closest to ready.
How it went
Cowork found 46 draft files, ran 44 searches against his site and suggested three drafts. One pick was a mistake: the Datasette upgrade guide already lived in the project docs and was never meant for the blog. It also built an unrequested animated page that was squeezed into a narrow sidebar.
Worth knowing
January 2026 first impressions of a research preview (Max plan, macOS). He warns that prompt injection is the main risk and that asking non-programmers to spot it is unfair.

simonw.substack.com ↗ · · Personal post

Full example with a prompt to try

Caiwei Chen (MIT Technology Review) · Manus

Put Manus to work on research lists and an apartment search

What they did
MIT Technology Review's reporter gave Manus three jobs: list notable journalists covering China tech, find two-bedroom New York apartments within criteria, and nominate 50 candidates for its Innovators Under 35 list.
How it went
The journalist list grew from 10 names to 30 after prodding and took just over an hour. The apartment search read "outdoor space" too literally at first. After three hours the 50-name nominee list was uneven. Crashes, "high service load" errors and blocked sites interrupted it; she put cost at about $2 per task.
Worth knowing
A March 2025 invite-only test that the reviewer called promising but not perfect; Manus's ownership and features have changed since, so check the current product.

technologyreview.com ↗ · · News report

Full example with a prompt to try

Related pages

FAQ

What can a general agent actually finish on its own?

Research with sources, a document or slide deck, a spreadsheet, a simple website, and multi-step browsing such as comparing options. It still needs you for logins, payments, judgement calls and a final check: treat the output as a first draft from a fast junior colleague.

How do credits work on Manus and Genspark?

Both charge credits per task rather than per message: a long research or website job can use hundreds. Manus lists Standard $20/mo for 4,000 credits plus 300 daily refresh credits; Genspark Plus is $24.99/mo from 10,000 credits. Check the credit balance after your first real task before you choose a plan.

I only need a repeating automation, not a general agent. What then?

A workflow tool is cheaper and more predictable for a job that runs the same way every time. See our rankings for your first AI agent and for building an agent without code.

Which of these are free?

Free plan confirmed on the official pricing page: ChatGPT, Manus, Claude, Genspark, Kortix.

Does anyone pay for these positions?

Rankings are built to avoid bias: no payment, commission or sponsorship changes a score or a position. Some links may become affiliate links (marked “Affiliate link”, rel="sponsored"). No affiliate links are active yet.