Last week, I hit my AI plan's limit in record time. The culprit was Fable, Claude's newest and smartest model.
Until now, I've always just picked the best model, because why not? Even Boris Cherny, Head of Claude Code, agrees:
"Always use the highest-performing model available… Using a cheaper model doesn't necessarily mean it's cheaper. In fact, using the highest-performing model often means fewer corrections and less hand-holding, so it gets the job done much faster."
But now Fable is double the cost of Opus (the next model down), and I couldn’t afford the always-pick-the-best rule anymore. I looked at the model and effort options in Claude’s dropdown and realized I actually had no idea how to choose. Would a lower model with higher effort be cheaper? Would the answers be just as good?
Here's what those settings are called:
The app | Its models | Its effort levels |
|---|---|---|
ChatGPT | The GPT-5 family (5.5 and 5.6), mostly hidden behind the effort labels | Auto (it picks for you), Instant, Medium, High, Extra High (Pro plans add more) |
Claude | Haiku, Sonnet, Opus, Fable | Low, Medium, High, Extra, Max, Ultracode |
Gemini | Gemini 3 Flash and Gemini 3 Pro (choosing Pro swaps the model, not the effort) | Fast and Thinking |
The method I use now comes from a guide Anthropic published for its programmers. When an answer disappoints you, ask: did the AI not know enough, or did it not try hard enough?
Not trying looks like skimming. You attach a twelve-page document and the answer only quotes the first three pages. You ask for five sections and get three tidy ones. For these cases, the fix is effort: turn the thinking up (or tell it, in plain words, to work through the full document before answering) and ask again.
Not knowing is harder to spot. The AI read everything and it followed your steps, but the answer is still confidently wrong, or the judgment feels mushy in a way you can't quite pin down. More effort won't fix this one; extra time gives the same brain longer to circle, but it doesn't make it smarter. This is a model problem. A bigger model was trained on more: more industries, more edge cases, more ways problems fit together, so it recognizes what you're asking with less spelling-out, and its judgment holds up on messier jobs. (It's also why the big ones cost more to run.) The fix: move up one model, or up the list if your app hides models behind settings, and run the same request again.
And sometimes no setting is to blame, because the AI never had the information it needed in the first place. If it guessed at your prices, your dates, or what your client said, that's not a model problem or an effort problem. It's missing information, and Issue 7 covered that fix: give it better files instead of better questions.
Don’t like the AI’s answer?
It skipped or skimmed → same model, more thinking.
It read everything and still got it wrong → bigger model.
It made up facts only you know → missing information.
Can't tell? → re-ask one step up and compare.
So, back to the question that started this: would a lower model with higher effort be cheaper, and would the answers be just as good? The answer: if the job just needs more time to complete (read everything, follow every step), then yes: a smaller model with the effort turned up costs less, and the answer holds. If the job needs judgment the smaller brain doesn't have, then no: effort can't make it smarter, so go straight to the bigger model.
The 30-second version
The model decides which brain you get, and the effort is how long it thinks.
A bad answer has two causes. Skipped steps and skimming mean it didn't try: turn up the effort. Confidently wrong after reading everything means it didn't know: move up a model.
If the AI guessed at facts only you know, no setting helps. It needed better information, not a better brain.
A smaller model with more effort is the cheaper pick for jobs that just need the hours. It can't rescue a job that needs judgment; there, the redo costs more than the big model would have.
Everyday work on the top model burns credits you didn't need to burn.
The settings earn attention twice: before a job you can't afford to redo, and after an answer disappoints.
Fable's price tag did me a favor. It turned that little dropdown from something I took for granted into a conscious decision.
me+machine.