The formula never changes
Imagine.
Nobody in these posts has used Opus 5 for more than an hour.
Every one of these posts follows the same three beats.
First, a bold claim that you are bad at this.
Second, a “secret” that turns out to be a system prompt trick everyone already knew about in the Sonnet 5 era.
Third, a call to action to buy a $19.99 prompt pack that will fix your obviously broken life.
You can tell because the screenshots are still full of the model correcting typos in its own demo prompts lmao.
My favorite specific claims
- One post said “99% of people don’t know Opus 5 can toggle effort levels.”
This was announced in Anthropic’s own launch post.
It is not a secret.
It is a feature.
Knowing a feature exists because the company that built it told you is not an insight, it is reading comprehension.
2. Another said the real trick is “talking to Opus 5 like a person, not a search engine.”
This has been the advice for every model since GPT-3.
At this point it is less a tip and more a genre convention, like starting a fairy tale with “once upon a time.”
3. A third claimed 99% of people are “wasting tokens” by not using the low effort setting for simple tasks, then spent 1,200 words explaining a toggle that has a label right next to it in the UI.
Why this keeps working
I run two Medium publication myself and edit for four, so I get why people write like this.
A headline that says “here is a small update to how you prompt” gets ignored.
A headline that says “you are doing everything wrong” gets clicked, because nobody wants to be the 99%.
But there is a cost to this.
Every one of these posts trains readers to expect a hidden trick, when the actual skill with any model, Opus 5 included, is boring.
Be specific.
Give context.
Iterate.
Read the actual docs the company published on release day instead of a hot take written forty minutes after the announcement.
Just do this for now!
What actually changed with Opus 5
If you want the non-clickbait version: it is priced the same as 4.8, gets close to Fable 5 on coding and knowledge work, and adds an effort toggle so you are not burning output tokens on a one-line answer.
That is it. That is the whole post for now.
Nobody can share tips and tricks until now.
Give it atleast 2–3 days, people.
Honestly,
I could have made this 200 words shorter and it would not have hurt anyone’s understanding, unlike the ten articles I just sat through.
The real 99% problem
The actual thing 99% of people are doing wrong is reading model launch takes written by people who had less hands-on time with the model than the average person spends deciding what to have for lunch.
Even this one, tbh.
Give it a week. Then come here and read me who might has actually shipped something with it (no guarantee).
I will let you know how the token toggle holds up on real workloads, once I have used it for longer than the time it took to write the headline.
My bucket of bets!