You paste 12 files into ChatGPT, ask for a refactor, then copy the answers back one by one. Or you spend a week wiring up the OpenAI API for a report you only need once.
The same models sit behind very different tools. Each one suits a different kind of work, and picking the wrong one wastes more time than it saves.
| Tool | Examples | Who drives | Where it works | Best for |
|---|---|---|---|---|
| AI chat | ChatGPT, Claude, Microsoft 365 Copilot | You, message by message | A chat window | Thinking, writing, and once-off questions |
| Coding agent | Codex, Claude Code | You set the task, the agent edits and runs code | Your repo, in a terminal, IDE, or cloud sandbox | Changing code and files |
| Always-on agent | Dots, Grok Bot | The agent works toward a goal in the background | Its own cloud computer | Ongoing work across apps and websites |
| API | OpenAI API, Claude API | Your code | Your backend | Product features and automation |
Start at the top and move down only when the tool above can't do the job.
Use AI chat when a person wants to work with AI directly and read every answer.
Good examples:
✅ Best for human-in-the-loop work.
Custom GPTs used to be the way to package a repeatable assistant inside ChatGPT. OpenAI stops creation of new custom GPTs on 25 September 2026 and retires them on 11 December 2026. Move them to ChatGPT plugins, or to a project with saved instructions.
Use a coding agent when the work lives in files. The agent reads your repo, makes the change, runs the build and tests, and shows you a diff.
Good examples:
Pasting 12 files into a chat window, asking for a refactor, and copying each answer back by hand.
❌ Figure: Bad example - Bad example - Using AI chat for work that lives in a repo
Asking Claude Code or Codex to do the refactor in the repo, then reviewing the diff and the test results.
✅ Figure: Good example - Good example - Using a coding agent where the code lives
✅ Best for changes to code and files, with a person reviewing the result.
See Do you use AI CLI tools? for how to choose between them.
Use an always-on agent when the work runs over hours or days, and you want to hand it off rather than sit with it. These agents run on their own cloud computers, so they keep working after you close your laptop.
Good examples:
Both products are new, so treat them like a new team member. Give each agent only the accounts and permissions it needs, and check its work before you rely on it.
✅ Best for ongoing background work across apps.
Use an API when AI needs to be part of a system rather than a conversation.
Good examples:
✅ Best for production systems, automation, and programmatic integration.
When an API feature grows into an agent that plans and calls tools, see Do you build agentic AI? To pick the model behind it, see Do you pick the best Large Language Model for your project?