I owe you this one.

In June I gave you a simple rule for when to splurge on Claude's smartest model, in July I told you it had broken, and then two of the things it depends on changed again, which is why the version below is worth more today than it would have been three weeks ago.

The piece with a deadline goes first.

1
Run /usage today, not Thursday.

Since May 13, Claude Code subscribers have been running on a temporary 50% boost to their weekly usage limit. Anthropic has it scheduled to end tomorrow night, Wednesday August 19, at 11:59 PM Pacific. If that holds, your week gets roughly a third shorter overnight.

I would put real weight on "if that holds." Anthropic has already extended this promo three times and could extend it a fourth. So don't plan around the date, check your own account. Type /usage into Claude Code today, note what it says, then look again on Thursday. That is your source of truth and it takes about ten seconds.

This doesn't touch everything, so don't over-panic. It applies to the weekly limit, not the 5-hour session limit, and the doubling of that session window is permanent. It's also Claude Code specifically. The separate Cowork boost ended back on August 5.

2
The most expensive setting is often the wrong one.

This is the part that broke the June rule.

Anthropic's own coding benchmarks show Opus 5 accuracy peaking at the medium effort setting, around 53%, then falling off as you push to high or max. You pay more and score lower. Independent testing from Vals AI found high effort did win on other kinds of tasks, so this is not a blanket "always pick medium." The point is narrower: the top setting costs more, does not reliably buy you more, and the interface is built to make you reach for it anyway.

There is a quieter cost underneath that. Tokenizer changes mean the same effort label now burns roughly 30% more of your allowance than it did before. You would never spot that from the interface, because the setting is named exactly the same.

3
Stop saving the good model for later.

The old instinct was to ration. Hold the expensive model back for the hard problems, grind through everything else on the cheap one. On August 10, Anthropic cancelled the 50% price increase it had scheduled for September 1, which makes Sonnet 5's introductory rate permanent.

The cheap model is now permanently cheap. Rationing stops making sense when the thing you are saving got a permanent discount, and the thing you are saving it for gets less accurate the harder you push it.

4
Know what actually moves your bill.

Opus 5's list price did not go up. What changes what you pay is the effort setting, whether fast mode is on (double the standard rate), and how much quota you have left. All three are under your control. None of them is the sticker price everyone argues about.

The revised rule

Default to the mid setting. Reserve the top setting for genuinely open-ended problems, the kind where you would take three swings at it anyway. Use the cheap model without guilt, because it is cheap on purpose now. And check /usage before you assume you have room.

That is the whole rule. It fits on a sticky note, which is roughly how I know it beats the June version.

Next issue: OpenAI shipped a desktop feature yesterday that logs your keystrokes to an unencrypted plain text file on your Mac. You ran the four checks on screen recording last month. This one needs its own.

If someone you know is paying for Claude and has no idea their limits change tomorrow night, forward this today. On Thursday it is just news.