Key takeaways
A reader who’d just used the model picker asked the next question. Which AI agents would you actually use?
The picker compares the models an agent is built on. It says nothing about the agents you can buy off the shelf: Manus, Devin, Operator, and the agents inside ChatGPT and Gemini.
In 2026 “agent” means whatever its seller needs it to mean. It can be a long-running autonomous worker, a chat with web search or a wrapper round a browser. The seven below all earn the name. They do different jobs, so this piece matches agent to job.
How I’d test them
I haven’t run these tests. These are the six tasks I would run, because they’re what people try first:
Run each agent through the same list. Count how often it finishes without help.
None are reliable enough to run unsupervised in production yet. Every pick assumes you’re watching what it does.
General-purpose agents
GPT Agent Mode, at $20 a month with ChatGPT Plus, is the place to start. It does most of what the others do, inside a more careful sandbox, for by far the lowest price. I’d expect it to be strongest on bounded tasks.
Manus, at $39 a month, is the step up. Check its claims on multi-step research and structured output against your own task. Read every result before you act on it, because it works in a cloud sandbox you can’t see into.
Gemini Workspace Agent, a $15 add-on to Workspace Business, is narrower. Think drafting the email with the prep doc attached, booking a one-to-one with everyone on a list, or summarising the last 30 emails from a client. If your team lives in Gmail, Docs and Calendar, it may be the most useful agent here.
Coding agents
Devin, at $500 a month for a team, is built to finish a software task with little supervision. Try it on one small, low-stakes task before paying. Read every pull request it opens. The cost per attempt climbs fast when half the attempts need a rewrite.
Its real use is a backlog of well-scoped chores, like dependency upgrades, simple endpoints and gaps in test coverage. Jobs you want done without a senior engineer’s time.
Replit Agent, at $25 a month, is the quickest way to build a small working app in one sitting, say a throwaway expense tracker. What it makes is a sketch you wouldn’t ship.
Computer-use agents
OpenAI Operator needs ChatGPT Pro at $200 a month. It’s the standard browser agent. It works in a cloud browser, sees the screen, clicks, types and fills forms. It can’t log in as you, so test it on a public-web task first. Only pay for it if you have a web task that repeats.
Claude Computer Use is priced through the API. It’s the only one that runs on a machine you control. That means setting up the machine, giving it the logins it needs and trusting it to do what you asked.
In return it can use any desktop software, local file and internal system. It’s the right choice for desktop work such as financial models in Excel. For most other teams it’s too much setup.
How do I choose?
Start with the task:
Then the budget. Under $30 a seat means GPT Agent Mode or Gemini. $30 to $50 means Manus or Replit. $200 or more buys Operator or, for coding teams, Devin.
All seven need watching today. Use them on tasks you were going to check anyway.
Pick one per task. Bouncing a job from Manus to Operator and on to a third tool usually ends with three half-finished attempts and three invoices.
What none of them do yet
None of them really remembers last week’s work unless you give it the context again.
Anything behind SMS codes or a hardware key fails. That explains a lot of “the agent couldn’t finish”.
Tasks that juggle three or more browser tabs at once fall apart.
“Figure out the best approach” gets you something, usually weak. Agents follow instructions. They don’t yet set useful goals.
None of them listens for events as they happen. Each runs once and reports back.
For jobs like these, a scheduled script, a tool like Zapier or your own hands will do better for now.
Building your own agent? The live model cost calculator compares per-token prices across providers. The frontier isn’t just American models covers the models underneath.
Three-minute lesson on this Workflow or agent? Marketing course, advanced. Free, no sign-up.