Fable 5 is arguably the best model out there (though Sol might have something to say about that). It is also the easiest model I have ever burned through. I kept handing it work that a cheaper model could have finished, watched my usage evaporate, and then did it again the next day. So I finally sat down and worked out how to actually rein it in. Five things made almost all the difference, and none of them require you to give up the good model.
Let Fable plan, and let cheaper models build
The most expensive habit is treating Fable as a laborer. It is at its best when it is breaking a big job into steps and deciding what happens in what order. That is the part worth paying frontier prices for. The actual building underneath, the file edits, the boilerplate, the repetitive passes, does not need that much horsepower.
Pick who does the building
So split the job. Let Fable write the plan and the architecture, then have it hand execution down to something cheaper. If you live inside Claude, point that execution at Opus 4.8. If you are like me and cheerfully pay for several subscriptions at once, you can point it at GPT-5.6 instead, which stretches your usage further and sits very close to Fable in overall intelligence.
Set the roles once
The trick is telling it upfront who does what. A few lines in your instructions file (your CLAUDE.md) that rank your models and say plainly which one plans and which one executes will get Fable routing the work itself, instead of doing every part of it the expensive way. Once those lines exist, you stop making the decision manually on every task.
I keep my own Fable orchestration skill for exactly this, and I update it constantly. It is a set of rules that tells Fable how to plan a job and fan the pieces out to the cheaper models underneath. It saves me from re-explaining the same delegation logic every session.
Dial the effort down
Fable runs at high effort by default. That is Anthropic's own recommendation in their prompting guide, and it is the right default for hard work. It is a terrible default for the routine stuff most of us do most of the day.
Drop to medium for anything ordinary and you will barely register the difference in quality while spending a fraction of the budget. Anthropic backs this up directly: lower effort settings on Fable 5 still perform well, and often beat what xhigh produced on the previous generation of models. The floor moved. Your habits from six months ago are now costing you money for nothing.
Skip the usage-hungry modes
Save high for the genuinely hard, multi-step problems, and reserve xhigh for work where capability is the whole point. Then leave the usage-hungry modes alone. Ultra, loop mode, and deep research all spin up swarms of sub-agents and tear through your limits fast. They are worth it occasionally. They are not worth having on by accident.
Keep your context lean
Don't let it run to the limit
The million-token window is a trap. It is so large that nothing stops you, so the natural behavior is to let a session run until it auto-compacts on its own. I compact well before that now, and it changed my usage more than any other single habit. A conversation that runs forever re-sends everything in it, and a bloated context drags the model's performance down on top of the cost.
Work one task per session, and open a fresh one when you switch jobs.
Hand off cleanly between sessions
When you genuinely need to carry work forward, have the current agent write the handoff before you leave. I say something close to this:
"I need to bring a continuation of this to another agent. Go through our whole conversation, pull out every insight and decision, and put together a detailed handoff document I can give another agent so it can start from scratch with all the context."
Then I open a new Fable session and tell it to use the handoff doc. The new session starts with the substance and none of the debris.
See where the tokens actually go
Two commands help you see where the money is going. Run /usage and scroll down. It breaks your consumption out by skill, subagent, plugin, and MCP server, each as a percentage of the total, which usually reveals something you forgot you installed. Boris Cherny, who leads Claude Code, says that when someone is mysteriously burning tokens, "most often, it's a bad plugin that's causing the issue." Run /context when you want to see what is filling the window right now, and /mcp to switch off servers you are not using.
You can also steer compaction rather than letting it guess. Saying /compact Focus on code samples and API usage keeps what you actually care about and drops the rest.
Watch the sticky defaults
Your model choice carries between conversations. Set one chat to Fable, open a brand new one, and it opens on Fable too. I have caught Fable grinding away on a trivial task more than once because I picked it for something big and a fresh conversation inherited that setting hours later.
Effort level behaves the same way. Bump one task to xhigh and every new conversation after it opens at xhigh until you go back and change it. Neither of these announces itself. You just pay for them.
So make it a habit: before starting a new task, look at which model and which effort level you are actually on, and drop them when the work does not need the firepower.
Tighten your prompts
Cut the pleasantries, and try not to rant at Fable. Every round of back-and-forth re-sends the whole conversation, and it adds up faster than it feels like it should.
When I do need to rant, I rant at Opus, then tell it: "okay, now give me a clean prompt to hand to Fable." The cheap model absorbs the mess and Fable receives the tight version. At the start of a session, set the terms in one line: "answer in five bullets, don't be verbose."
Say less
Keep it short. Anthropic found that Fable follows instructions well enough now that a single short brevity line steers it about as well as an enumerated list of every behavior you want. Everything past that first line is tokens you are paying for.
That same logic applies to the rules and skills you already wrote, so you may want to audit them too. Anthropic says skills developed for prior models are often too prescriptive for Fable 5 and can degrade output quality. The elaborate scaffolding you built to make an older model behave may now be working against you.
Start with the effort setting and the sticky defaults. Those two take about a minute and account for most of the waste.