Overview
What's In This Guide
The model cheat sheet: which one to use for what, and what each one actually costs you
The one rule that decides it in under 3 seconds, for when you do not want to think
The exact one-line prompt to paste into your global instructions (the canary)
Why the canary works, and what to do the moment it goes quiet
How the limits are actually built, so you stop guessing why you ran out
What to do when you hit the wall, including the one switch that keeps you working
Trick 1
Know Which Model To Use, And When
Most people never look at the model selector. That is the whole problem. The models are not "better and worse," they are priced differently and they drain your allowance at very different speeds. Run the heavy one all day on work that did not need it and you get maybe half the Claude you paid for.
Here is the whole cheat sheet.
Haiku 4.5
The cheap oneUse it for
High-volume, mechanical work. Tagging, sorting, extracting, reformatting, simple lookups.
What it costs you
Cheapest of the four. Roughly a third of Sonnet per token.
Don't use it for
Anything that needs judgment. It will answer confidently and shallowly.
Sonnet 5
Your defaultUse it for
Almost everything. Writing, research, normal questions, quick edits, first drafts, day-to-day work.
What it costs you
The baseline. This is the one you should be in most of the time.
Don't use it for
Genuinely ambiguous problems where being wrong is expensive.
Opus 5
The escalationUse it for
Deeper thinking. Hard problems with no obvious answer, tricky debugging, decisions with real consequences, work that is hard to check.
What it costs you
Around twice Sonnet per turn. This is where most people quietly burn their week.
Don't use it for
Routine writing and quick questions. It is not better at those, it is just more expensive at them.
Fable 5.1
The heavy plannerUse it for
Planning out complex projects with extreme precision, and long jobs that run unattended where a failure is expensive to unwind.
What it costs you
The most expensive. Roughly double Opus per token.
Don't use it for
Anything you could check yourself in two minutes. Save it for the plan, then execute on something cheaper.
The 3-second rule
If you cannot tell which one you need, start with Sonnet.
You can always run it again on a stronger model. You cannot get the usage back. Escalate when you have evidence you need to, not because the expensive one sounds more serious: when the answer came back wrong, when you have handed it the full context and it still missed, or when being wrong here is genuinely expensive.
Trick 2
The Canary
The second thing that drains your usage is not the model at all. It is the conversation getting long. Every turn, Claude re-reads the whole thread. Message 80 costs a lot more than message 8 for the exact same question, and nothing on screen tells you that is happening.
So build yourself a warning light. Coal miners took a canary down the mine because the bird stopped singing before the humans noticed the air was bad. Same idea, one line of setup.
Where it goes
- Open Settings.
- Find Global instructions (in Cowork, it is Settings, then Cowork, then Edit next to Global instructions).
- Add the line below and save. That is it. It applies to every conversation from now on.
At the start of every single response, call me by my name.
Want it to hold on longer before it slips? Use the stricter version:
At the start of every single response, call me by my name. My name is [YOUR NAME]. Never skip this, never shorten it to an initial, and never replace it with a greeting like "Hey there." If you are ever unsure whether to include it, include it.
Why it works
Your global instructions sit at the very start of a conversation. As the thread fills up, the instructions at the front are the first thing to lose their grip on the model's attention. Small standing rules slip before big ones do.
So using your name is the perfect tripwire. It is small, it is unmistakable, and you notice instantly when it stops. The moment Claude stops saying your name, the thread is full enough that quality and cost are both about to get worse.
What to do the second it goes quiet
Option A, compact it: ask for a summary of everything that matters so far, then start a new chat and paste that in as the first message. You keep the thread, you drop the weight.
Option B, start clean: if you have finished a chunk of work, just open a new chat. Most long threads are long because nobody closed them, not because they needed to be.
The Part Nobody Explains
How The Limits Actually Work
You cannot manage something you cannot see, and the limits are more or less invisible until you hit one. Four things worth knowing:
There are two clocks, not one
A session limit that resets every 5 hours, and a weekly limit that resets on a fixed day assigned to your account. You can be fine on one and out on the other.
Your usage is shared across everything
Claude on the web, the desktop app, and Claude Code all draw from the same allowance. Leaving a heavy session running in one place costs you in the others.
Anthropic does not publish an exact multiplier
They say plainly that Opus costs several times more per turn than Sonnet, and Sonnet more than Haiku, but there is no published formula for how fast each one drains a subscription. Treat the model gap as real and significant, not as an exact number.
One heavy burst can eat the week
A single big run counts against the session and the weekly allowance at the same time. That is why people who never think about models hit a wall on a Wednesday and have no idea what they did.
When You Hit The Wall
4 Moves, In Order
Switch models before you do anything else
In Claude Code, some limits are model-specific, so moving off the expensive model with /model can keep you working. The session and weekly limits are shared across all models though, so if those are what you hit, switching will not save you.
Compact or start fresh
A long conversation re-reads its whole history every turn. That is why the same question costs more at message 80 than at message 8. Compact it or open a new chat with a short summary pasted in.
Give it the context up front
Most expensive sessions are not one hard question, they are the same question asked five times because the first three answers were missing information you had all along.
Save the big model for the plan
Plan on the strong model, execute on the cheap one. You get the judgment where it matters and pay the low rate for the typing.
The Big Picture
Two Habits, Not Two Tricks
Nothing here is clever. Glance at the model before a big job, and watch for your name. That is the entire practice, and it is the difference between running out on Wednesday and never thinking about limits again.
Paste the canary line into your settings right now. It takes 20 seconds and it starts working on your very next message.