Models, effort and limits
The model is which Claude you ask. Effort is how hard that model works on this one reply. Limits are how much of your plan you've used. Get those three wrong and a simple email costs as much as a board pack.
How to access it
The model picker is in the chat, in the desktop app and on the web. Effort sits with it. Usage is under Settings > Usage.
Paid plans (Pro and up) are what we set up in the room. Check the current price at claude.com/pricing. Fable draws on usage credits faster than the others. A Pro plan includes Opus. Fable is there via those credits, and it runs the tank down quickly.
The switch people miss: effort. Medium is the default. High is for a one-off piece of work that has to be right. Ignore Max, and ignore Extra high if you see it. You don't need it for a normal day's work.
Which model
Think of it like cars. Sonnet is the one you drive every day. Opus is the one you take when the trip has to go well. Fable is the fastest, and it uses the tank quickly, so you save it for the work that really matters.
| Model | Use it for |
|---|---|
| Sonnet 5.5 | Everyday drafting, mail, summaries, the morning brief, most skills |
| Opus 5.5 | One-off high-value work: tone of voice from your sent mail, a contract first pass, a model, a board pack |
| Fable 5.1 | The rare job where being wrong is expensive. Expect it to eat limits |
| Haiku 5.5 | Short, cheap tasks when you can see it in the picker. Not the default we teach |
Older names (Sonnet 5, Opus 5, Fable 5) may still appear. Prefer the current ones above.
Don't switch model in the middle of a chat. Claude re-reads the conversation, and a switch throws away the saving that makes that re-read cheap. If you need a different model, ask for a 10-line summary and start a new chat with that summary pasted in.
Effort
| Effort | When |
|---|---|
| Low | A tidy, a translation, a grammar pass |
| Medium | The default. Leave it here |
| High | One important piece of work. Then go back to Medium |
| Max | Ignore it |
High on every message is how people burn a week of usage on Tuesday.
Why you hit the limit
Every new message re-reads the whole chat. Then it thinks (that's effort). Then it writes. A long thread, a heavy model, files attached, web search, and several connectors all count.
Limits reset on a rolling window of about five hours, and paid plans also have weekly limits. Both draw on the same pool whether you're on the web, the desktop app, or the phone. Settings > Usage shows where you are.
On Team, a Premium seat has more usage per five hours than a Standard seat (the pricing page describes Premium as 5x Standard).
What to do when you're burning it:
- Don't revive a huge old chat. Ask for a 10-line summary, then start fresh.
- For a quick question: "Answer here. No web search. No files. Under 150 words."
- Plan a deck or a long doc in text. When the outline is right: "Build this exact outline."
- Turn off connectors you aren't using. Each one is loaded in.
- Highlight the bit that's wrong and edit that, if the app offers it, instead of redoing the whole answer.
- A daily task on Opus or Fable will hurt. Sonnet for the schedule. Opus for the job that's worth it.
Summarise this chat in 10 lines: decisions made, facts I gave you, what's still open.
I'll paste that into a new chat. Don't add anything we didn't cover.
Worked example. Someone kept a single chat for a month of client work, then switched to Fable halfway through "to make it better". The next reply was slow, the usage bar jumped, and the answer repeated context they no longer needed. They took a 10-line summary, opened a new chat on Sonnet at Medium, and finished the email. Fable stayed for a contract pass the week after, in its own chat.
Plans
Prices on claude.com/pricing were in dollars when we last checked, and the UK pound figure wasn't on that page. Don't budget from a converted number. Check the price while you're logged in, or on the UK view of that page.
| Plan | What it's for | What's included that we use |
|---|---|---|
| Free | A look, not a session | Chat, files, memory, connectors, a small number of projects. Design is listed. Not enough usage for real work |
| Pro | One person | More usage, scheduled tasks, Dashboards (beta), Claude Code, Claude in Chrome, Research, projects. Working in a folder, in the same app as chat. Opus included. Fable via usage credits |
| Max | Heavy individual use | More usage than Pro (5x or 20x), higher output limits, early access. Still not a requirement for working in a folder |
| Team | A company, 2 to 150 people | Standard and Premium seats. Shared projects, org-wide skills, connector admin, org instructions, Motion (beta). Training off by default |
| Enterprise | Larger organisations | Team's features, plus SCIM, audit logs, custom retention. Priced as a seat fee plus usage |
Working in a folder is on the paid plans, including Pro and Team. It is the same app as chat. Max is a usage tier.
Creating a Team can pull in an existing account on that email and cancel Pro or Max. Skills and connectors don't come across. Tick keep-personal-account-separate if you still need it.
Pro tips
- Sonnet, Medium, new chat when the thread gets long. That's the whole habit.
- Opus for the tone-of-voice file, once. Not for every email after that.
- If a reply is wandering, lower the effort or tighten the prompt. Don't reach for a bigger model first.
Limits of this page
The picker labels move. If you see Extra high, or a model name that isn't in the table, the table is behind the product. Use the rule anyway: everyday model, a heavier one for a one-off, the top one rarely.
Anthropic can also throttle features when it's busy. If a model vanishes for an hour, it isn't your folder.
Where to start in 30 minutes
- Open Settings > Usage so you know the baseline.
- Set the next chat to Sonnet, effort Medium.
- Do one real email in that chat.
- Open a second chat on Opus only if the job is the tone file, a contract, or a pack. Close it when you're done.
- Add the "no web search, no files" line to your notes for quick questions, so you don't spend a research budget on a yes or no.