Cowork vs ChatGPT Work vs Manus: Price the Task You Will Accept
August 2026 put three flavours of knowledge-work agent in the same conversation: Anthropic Cowork, OpenAI’s Work-style agents, and credit-based products like Manus. They are not substitutes.
This page is periodically reviewed to reflect current pricing and plan changes.
Fastest win
Pick the agent that matches the artefact. Local files: Cowork, with backups. Connected office work: ChatGPT Work or the Microsoft/Google equivalent. Open-ended multi-deliverable jobs: a credit agent like Manus, metered hard. Do not buy all three.
Three products, three units
Cowork, as reported on 20 August 2026, is a local-folder agent on Claude Max / macOS. ChatGPT Work-style products meter long-running office tasks against plan capacity and extra credits. Manus-class products sell credits against delegated multi-step jobs.
If you compare them on “intelligence”, you will buy the wrong one. Compare them on the artefact: a cleaned folder, a spreadsheet in a connected drive, a deployed mini-site.
Our existing comparison pages for ChatGPT Work vs Claude Cowork and Manus vs Genspark still apply. The news is that Cowork is now a concrete preview, not a rumour.
Permissions are part of the price
A cheap credit that can delete files is not cheap. A Max seat that cannot leave a selected folder may be the bargain.
Write the permission boundary first. Then price the vendor who can live inside it. After the month’s containment stories, this is not optional colour. It is the product.
Cowork or Claude for file-heavy synthesis; do not default to credits
Credit-based general agents look flexible and then eat the month. If the job is a folder of documents, a desktop agent on a plan you already understand is easier to govern.
Cost per accepted completed task
Numerator: plan + credits + retries + reviewer time + undo time. Denominator: tasks a human marked acceptable.
Five representative tasks will tell you more than a launch video. Include one failure on purpose. You need to know what a bad run costs.
Ranked recommendation
Best choice for local document assembly on a Mac: Cowork if you were going to buy Max anyway and you have backups.
Best choice for connected office artefacts: ChatGPT Work or the suite-native agent, not a third credit pack.
Best choice for open-ended delegated jobs: a credit agent with a cap, not boundless enthusiasm.
Avoid stacking Cowork, Work and Manus on one person in the name of evaluation.
Key Takeaways
- →Artefact type picks the agent family.
- →Credits, seats and included runs need one denominator: accepted tasks.
- →Permissions and undo time belong in the numerator.
- →Do not run three general agents as a hobby.
Editorial context
Who is this for?
Founders and ops leads shopping “an agent” after Cowork’s preview.
When NOT to use this
Developers choosing coding agents. That is a different accepted outcome — merged code — and a different page.
Pricing insights
Subscriptions, credits and “included agent runs” are not comparable until you share a task definition. Failed runs count.
Alternatives to consider
Perplexity Computer for cited web work. n8n or Make if the job is actually a workflow with an AI step. A human plus Claude Pro if volume is low.
Final verdict
One agent family per artefact type. Measure accepted tasks for two weeks before you add a second family.
Frequently Asked Questions
Which general agent is cheapest in August 2026?
The one with the highest accepted-task rate after review, not the lowest credit sticker. Include retries, runtime and the human who undoes a bad run.
Should we buy Cowork, ChatGPT Work and Manus together?
No. Pick one bounded workflow, run five tasks, and retire the others unless a second job is named. Parallel pilots become parallel shelfware.
What changed after the safety reporting?
Permissions and logging became line items. An agent that cannot show what it touched is incomplete even if the demo looks fast.