You’ve probably noticed it. Your Claude Opus 5 — the model that’s supposed to be faster and cheaper — is suddenly burning through tokens like a teenager with a credit card. The dashboard looks fine. The output looks impressive. But your costs are spiking, and you can’t figure out why.
Here’s the dirty secret: that harmless little prompt you’ve been using for years — ‘please check carefully’ — is now a financial landmine.
Opus 5 is different. It’s not a junior developer who needs encouragement. It’s a hyper-competent senior engineer who, when told to ‘be careful,’ will triple-check every line of code, spawn sub-agents to verify its own work, and expand the scope of a five-minute task into a full-blown audit. The more you tell it to ‘make sure it’s right,’ the more it will find ways to prove it’s right — and charge you for every single one.
I saw a team waste 50,000 tokens in a single session because they left their old prompt template unchanged. The model wasn’t broken. The prompt was.
The Paradox of Power
Here’s the uncomfortable truth: the more capable your AI, the more rigid your guardrails need to be. Opus 5 is a beast. It can handle complex multi-file refactoring, end-to-end code tasks, visual replication, and self-verification. But that power comes with a price: it will default to over-engineering every request unless you explicitly tell it to stop.
Think of it like giving a race car driver a faster engine. If you don’t install better brakes and tighter steering, you’re going to crash — or in this case, burn through your budget in record time.
Stop Saying ‘Check Carefully’
Your old prompt probably ended with something like: ‘Please check carefully and make sure there are no issues.’ With Opus 5, that’s not a safety net. It’s a blank check.
Instead, write your acceptance criteria as a cage. Be specific about what ‘done’ looks like:
- Not ‘verify everything’ — but ‘verify only the three critical paths’
- Not ‘identify potential issues’ — but ‘only flag issues that affect runtime performance’
- Not ‘improve the code’ — but ‘only change lines that have a bug, not style’
Your job isn’t to motivate the AI. It’s to limit it.
The Code Review Trap
When Opus 5 reviews code, it has a natural tendency to be too conservative. If you ask it to ‘be careful not to miss anything,’ it will flag everything — including stylistic preferences that aren’t bugs. The result: a firehose of false positives that your team has to wade through.
The fix is counterintuitive: let it be aggressive in the first pass, then filter. Tell it to report every single suspicious line. Then, on the second pass, ask it to rank by severity. Let the human do the final call. This way, you leverage the model’s strength (coverage) without being overwhelmed by its noise.
Sub-Agents: The Silent Cost Multiplier
Opus 5 loves to spawn sub-agents. For complex tasks, that’s a feature. For a simple button-text change, it’s a disaster. I’ve seen it spin up three parallel agents to analyze a single line modification — tripling the cost for no benefit.
Set hard rules: no sub-agents for tasks under a certain complexity threshold. Write it into your project-level instructions. Don’t trust the model to manage its own resource usage — it will always err on the side of over-engineering.
The Visual Task Hack
If you’re using Opus 5 for front-end replication or chart analysis, stop describing the layout in words. Give it tools. Give it a screenshot, a pixel comparison tool, and a DOM inspector. The model’s ability to ‘see’ is wasted if you’re just describing what you want it to see. Feed it real feedback loops, not vague descriptions.
The Bottom Line
Opus 5 is a game-changer — but only if you change your game. The teams that succeed will be the ones that stop treating it like a clever assistant and start treating it like a powerful, expensive engine that needs hard limits. Your prompt is no longer a request. It’s a contract. Write it like one.
If you keep using the same prompts from six months ago, you’re not just wasting money. You’re actively sabotaging your own workflow. The model has evolved. Have you?
FAQ
Q: Why does 'please check carefully' cause problems with Opus 5?
A: Opus 5 has a built-in tendency to self-verify and expand scope. When you add an open-ended request to 'check carefully,' it interprets that as a license to perform exhaustive verification, spawning sub-agents, re-reading files, and questioning its own output — all of which consume tokens exponentially.
Q: What's the practical way to constrain a powerful AI like Opus 5?
A: Write acceptance criteria as hard boundaries. Instead of 'verify everything,' specify exactly what to check and when to stop. Use conditions like 'only flag issues that break the build' or 'limit verification to the three critical paths.' Set project-level rules on sub-agent usage and effort levels.
Q: Is Opus 5 actually cheaper than Fable 5?
A: Yes, on a per-token basis it's about half the price of Fable 5. But if you use the same prompts and workflows, your total cost can actually be higher because Opus 5 tends to generate more tokens and spawn more sub-agents. The savings only materialize if you optimize your prompts and guardrails.