The fastest way to under-use AI in 2026 is to treat “Claude” and “Claude Code” as the same thing. They aren't. They are different products, with different cost dynamics, different failure modes, and different right-uses. The same applies on the OpenAI side with Codex CLI, and on the open-source side with Aider, Cline, and the rest of the field. Everyone in the room next month has used at least two of these. Almost nobody is using each of them for the work they're best at.
Claude in the chat box
The web app at claude.ai. The mobile app on your phone. Connectors that let it read your Drive, your Gmail, your Linear. This is the surface most people start with and stay on for too long. It is excellent at:
- Long-form writing where you want to iterate quickly with a smart reader.
- Reading long documents and giving you back an opinion.
- Project work where the same conversation accumulates context over days.
- Mobile-first work — quick triage of an email thread on the way to a meeting.
It is not good at — and was not designed for — writing code that touches your filesystem, running shell commands, or operating across more than a few connected services in one motion. If you find yourself copy-pasting code in and out of the chat window, you are using the wrong tool.
Claude Code
The CLI agent that runs in your terminal, on your machine, with access to your files, your git, your shell, and your tools of choice. This is what you should be reaching for the moment the work involves a codebase. It is the difference between “an assistant that drafts code for you to paste” and “an engineer that opens the file, makes the change, runs the tests, opens the PR, and reports back.”
The mental shift that matters: Claude Code is not a chat tool with file access. It is an autonomous agent in your shell. You give it a goal, you give it permissions, and you let it operate. The right rhythm is “here is the problem, here is the constraint, go,” followed by a code review, not a turn-by-turn conversation.
It also has the property — useful and underrated — that everything it does is logged in your terminal. The audit trail for an agent run is just your scrollback. That is not nothing.
Codex CLI and the OpenAI side
Codex CLI from OpenAI plays in the same space. We use both at IMPT.io for different work — Codex CLI tends to be slightly faster on small refactors; Claude Code tends to be more robust on architectural changes that need to be held in working memory across many files. Neither is a clean winner across the board. The pragmatic answer is that both should be available to your engineers, and the engineer should choose for the work in front of them.
Open-source agent framings
Aider, Cline, Continue, OpenHands. The open-source agent space matured in 2025 to the point where you can do most of what Claude Code or Codex CLI does on top of any model — including the Chinese open-source models we discuss in the next topic. The cost dynamics are very different: you pay for the model you're calling rather than the agent, which makes self-hosted setups dramatically cheaper at volume.
The trade-off is engineering effort. The polished commercial agents (Claude Code, Codex CLI) ship with sensible defaults, working tool integrations, and a UX that respects your time. The open-source agents require you to wire those up. For a small team doing a lot of agent-coded work, this is worth doing — the cost savings compound. For a team doing occasional agent-coded work, it isn't.
When to use each
| Work | Right tool |
|---|---|
| Drafting an email or a brief | Claude (chat) |
| Reading a 200-page legal document | Claude (chat) with extended thinking |
| Refactoring a Python module | Claude Code or Codex CLI |
| Building a new microservice | Claude Code |
| High-volume code-review across many PRs | Open-source agent on a self-hosted Chinese OS model |
| Production agent inside your product | Anthropic API or self-host, never the chat UI |
Cost in practice
Two things matter for cost. First: most of your spend should be on small specialised workers, not on the frontier model. A swarm where the planner and auditor use Claude Sonnet but the workers use a fine-tuned 7B-class model on a domestic GPU costs an order of magnitude less than a swarm where everything calls the frontier API. Second: prompt caching is the largest single lever. Anthropic's prompt cache pricing turns long-running agent loops from expensive to cheap if your prompts are structured to maximise cache reuse. We get our cache hit rate above 90% on most production swarms. That alone is the difference between a system that pays for itself and one that doesn't.
What we'll do in the workshop
Day 1 evening of the Clonmel workshop is hands-on. Everyone in the room with a laptop will install Claude Code, run a real refactor on a sample codebase, and then run the same refactor through an open-source agent on a Chinese open-source model — and compare. By the end of the session you'll know which tool is right for your work and you'll have it running on your own machine.