Opus 5.5: Initial Thoughts and Prompting Advice
Our first impressions of Claude Opus 5.5, and how to prompt it for the best results.
- ai
- claude code
- prompting
Opus 5.5 is out, and we have spent the last 12 hours experimenting with it.
In our experience, the improvements aren’t monumental. But they are tangible, and you will notice them within your first few prompts.
The biggest win for us is cost. We now get Fable-level code reviews for less than half the price. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, against $10 and $50 for Fable 5.1. It’s also quite faster than Fable.
Read on as we share our experience with Opus 5.5 so far, and how you can get the best out of it.
It Needs a Lot Less Hand-Holding
Opus 5.5 goes much deeper by default than Opus 5 did. In practice, this means you can hand it bigger tasks with less oversight and a much thinner spec.
We tested this with two versions of the same prompt.
The first was a detailed spec. It laid out our constraints, our preferences, and how we wanted the work done. The second skipped all of that and only described the task and the definition of done.
Both prompts produced largely similar output.
Anthropic’s early testers saw the same depth on much larger jobs. One of them audited and fixed a 200,000-line codebase in under three hours. Opus 5 took more than 20 hours on the same job.
Its Responses Are Finally Easy to Read
The second thing we noticed was how much easier its answers are to read.
One of the biggest complaints about recent Claude models has been their writing. The answers were usually correct, but the sentences were dense and the word choice was odd. You know what we are talking about, right?
Opus 5.5 has clearly had work done here. In our view, it’s the first model since Opus 4.6 whose default response style is easy to process.
Here’s the same question put to both models.
The prompt
In 2 paragraphs max, explain what leverage is in financial markets.
Opus 5.5
Leverage means you use borrowed money to control a larger position than your own cash allows. Say you have $10,000 and your broker lends you $40,000. You can now buy $50,000 of an asset. This is 5:1 leverage. You get leverage in several ways. You can borrow on margin from a broker. You can also trade derivatives such as futures, options, and CFDs, which need only a small deposit (the "margin") to control a large exposure. Companies and funds also use leverage when they borrow to finance their operations or investments.
Leverage makes gains and losses larger by the same factor. With 5:1 leverage, a 10% rise in the asset gives you a 50% gain on your own money. A 10% fall gives you a 50% loss. A 20% fall erases your whole deposit. If losses bring your account below the broker's minimum margin, you get a "margin call." You must then add cash, or the broker sells your position, often at a bad price. You also pay interest on the borrowed money. Some products can lose more than your deposit. So leverage increases your potential return and also your risk of fast, large losses.
Opus 5
Leverage is the use of borrowed money, or instruments that behave like borrowed money, to control a position larger than your own capital. A trader with $10,000 who borrows another $40,000 controls $50,000 of stock at 5:1 leverage. The borrowing can be explicit, such as a margin loan or a mortgage. It can also be embedded in the instrument: futures, options, and swaps require only a small deposit against a much larger notional exposure. Companies use the same idea when they fund operations with debt instead of equity.
Leverage multiplies percentage gains and losses on your own capital in equal measure. In the example above, a 10% rise in the stock returns 50% on the trader's $10,000; a 10% fall wipes out half of it. Two extra risks follow. First, interest and financing costs accrue whether the position moves or not. Second, lenders demand collateral, so a price move against you triggers a margin call, and forced selling can lock in the loss at the worst moment. This is why leveraged positions fail at price levels far short of the ones that would make an unleveraged position worthless.
Both answers are correct. But one is clearly easier to process than the other.
Look at how Opus 5 explains the risk. Its last sentence is “This is why leveraged positions fail at price levels far short of the ones that would make an unleveraged position worthless.”
It’s true, but if you don’t already know what leverage is, it’ll likely not make sense to you.
Opus 5.5 makes the same point with a number: “A 20% fall erases your whole deposit.” Its sentences are short and carry one idea each.
When it does use a term like “margin call,” it explains it right away.
This matters most when you ask about things you don’t know well. And if you’re a founder asking your vibe coding tool why something broke, that’s like… several times a day.
Stop Telling It to Think Hard
If you have been adding “think hard” or “think carefully before answering” to your prompts, it’s time to take it out.
Anthropic’s Opus 5.5 prompting guide recommends removing “think carefully” instructions from chat system prompts. In their testing, removing such a line made replies start sooner, with no clear decline in quality.
So the extra instruction mostly just makes you wait longer.
If you want more or less thinking, change the effort level instead. It’s the setting that controls how much the model thinks, and Anthropic says lowering it cuts thinking more reliably than prompt instructions do.
One thing to know if you use the API: Opus 5.5 defaults to medium effort, one level below the “high” default on Opus 5. If you still want to work with “high”, you can set it explicitly in the API call.
Read more: A Vibe Coder’s Guide to Prompting
Our New Favorite Model for Code Reviews
Opus 5.5 is now our go-to model for code reviews. We gave it this prompt:
Code review prompt
Compare the currently checked out feature branch with master.
If there are any issues that should block the merge, list them.
For each issue, briefly, in plain English, explain what the problem is, and how to mitigate.
The feedback was pointed and short. It listed all the problems, explained each one in a few plain sentences, and stopped.
It also didn’t invent feedback for feedback’s sake. (If you’ve asked an AI for a code review before, you know the kind)
Implementation Plans That Need Almost No Edits
The implementation plans Opus 5.5 wrote for us needed little to no changes.
Opus 5 had a habit of padding its plans. It would add steps nobody asked for, simply because they seemed good and relevant. None of it was exactly wrong, but you still had to find it and cut it before you could approve the plan.
Opus 5.5 sticks to what you asked for. The plan covers the task, and not much else.
If you’ve been adding lines like “don’t add anything I didn’t ask for” to your planning prompts, try a plan without them first. You may find you don’t need them anymore.
Videos!
Okay, now this one was unexpected.
Opus 5.5 can make a short video in one shot, without a video model, an external API, or any extra tools. It does this by writing the video as code.
We asked Claude Code for a simple 3-second title card to see how it works. It drew the animation on an HTML canvas with JavaScript, then used Chrome to play the animation and record it into a WebM video file.
The whole render took just five seconds.
Keep in mind that this is motion graphics: text, shapes, charts, and animation. It won’t produce realistic footage of people or places the way an AI video model does. But you will be able to create launch videos, product demos, or animated explainers.
And because the video is just code, you can ask for changes the same way you would with any other code. Slow down the intro, change the colors, or swap the text, and it renders again.
Here’s a three-line prompt to try:
Video prompt
Make a 15-second launch video for my app, [app name], at 1920x1080.
Open with the name, show these three features one at a time with smooth animations: [feature 1], [feature 2], [feature 3], and end on [your URL]. Use these brand colors: [colors].
A Security Review Prompt You Can Run Right Away
Finally, here’s a prompt for a security review of your entire codebase. We ran it with Opus 5.5 and got really good results, and we expect you will too.
Security review prompt
Do a security review of this entire codebase. Assume an attacker has our public URL and a free account, and wants to read or change data that isn't theirs, take over other accounts, or run up our bills. Check authentication and sessions, authorization on every API route and database query, input handling (injection, file uploads, requests to internal URLs), secrets in the code or git history, dependencies with known vulnerabilities, and the config for our database, storage, and hosting. Read the code before you judge it. Don't guess from file names.
Only report issues you can point to in the code. For each one, give the file and line, explain in plain English what an attacker could do and how, rate it critical, high, medium, or low, and describe the fix. List critical and high issues first. Don't pad the list with style advice or generic best practices. Don't change any code yet.
Read more: Is Your Vibe-Coded App a Security Risk? Here’s How to Find Out
Wrapping Up
Opus 5.5 isn’t a giant leap. But it’s cheaper, faster, easier to read, and needs less hand-holding. If Opus 5 is still your daily model, you should certainly switch.