By Dan Donovan, @askdandonovan
Four models, one lookup table, and the setting that stops you paying twice for the same document. Find your row and stop guessing.
Open your Claude settings and go to Usage.
You will see progress bars for how much of your five hour session and your weekly limit you have already used. Most people have never opened this screen, which is why hitting a limit always feels like it came out of nowhere.
Look now, before you change anything. Then look again in a week. Everything below is easier to believe once you can watch the number move.
Find your row. That is the whole decision.
| What you're doing | Use |
|---|---|
| Quick fact, tidy up an email, short rewrite | Haiku |
| Most work questions, writing, planning, research | Sonnet |
| Rewriting your CV against a specific job ad | Sonnet |
| Trip planning, meal plans, sorting out a budget | Sonnet |
| Drafting something in your own voice | Sonnet |
| Reading a long contract and pulling out what matters | Opus |
| Something that runs for a while without you watching | Opus |
| The single hardest piece of analysis you have | Fable |
If your task is not on the list, use Sonnet. That is not a cop out. Sonnet is the default for a reason, and it is strong enough that most people never genuinely need to leave it.
Picking a heavier model does not lock you out of anything. It spends your allowance faster. Anthropic is explicit that how much you can get done depends on the length and complexity of your conversations, the model you choose, and the features you use.
So the cost of always reaching for the biggest model is not that you get blocked. It is that you run out sooner, for an answer that often would have been the same.
That habit made sense when the cheaper models genuinely were not good enough. That gap has narrowed a lot this generation. Anthropic's own migration guidance now says the lower settings on the current models often match or beat what the previous generation produced at its highest settings.
Model choice is one lever. This is the one most people have never heard of.
Put anything you reuse into a Project. In Anthropic's own words:
So if you keep pasting the same brief, the same brand guidelines, the same CV, the same product details into fresh chats, you are paying for that content every single time. Upload it to a Project once and the reused part stops counting.
Do not take my word for any of this. Take a task you would normally hand to the biggest model, run it on Sonnet, and paste this on the end:
Before I accept this, be straight with me: did this task actually need the most capable model, or would a faster one have produced an answer just as good?
Point to the specific parts of your reasoning that needed the extra capability. If there aren't any, say so plainly.Do that three or four times on real work. You will learn where your own line sits faster than any comparison chart will tell you.