The previous three parts were about what Opus 5 can do. This part is about what breaks when you move to it. It is the narrowest piece in the series: relevant if you build on the API or run agents, skippable if you just use Claude.
Anthropic says Opus 5 performs well on existing Opus 4.8 prompts. That is true, and it is exactly why these things catch you out. It mostly works, right up until it does not.
Two things that throw an error
Thinking is on by default now. On Opus 4.8, leaving out the thinking parameter meant no reasoning. On Opus 5, leaving out the same parameter means adaptive thinking. No error appears, your bill just changes shape.
The side effect is worse. Your max_tokens is a hard ceiling on thinking plus response together. If you tuned it tightly around your expected answer length, the model can now be cut off mid-answer. Every route where you never set thinking deserves a look: switch it off explicitly, or give max_tokens room.
Disabling thinking is capped at high effort. Combining thinking: disabled with xhigh or max returns a 400. This is validated per request, so a later call that raises effort while thinking is still off will fail even though earlier calls succeeded. Walk your call sites, not just the first one.
My advice on both: leave thinking on and steer with effort instead. As part two covered, low and medium on this model do work that used to require high. That is a better cost dial than switching thinking off.
Two things that quietly cost you money
These throw no error. They make your system more expensive or worse, and you find out on the invoice or in the output quality.
Your verification instructions now work against you. If your prompts or harness tell the model to double-check its work, or to run a separate verification step, or to spin up a subagent to review: delete it. Opus 5 verifies itself unprompted. Leaving the instruction in produces duplicate work, more tokens, and longer chains without a better outcome.
This feels wrong, because “ask the model to check its work” has been sound advice for years and appears in nearly every prompting guide. On this model it inverts. If you maintain a prompt library, this is the rule that needs a carve-out.
Your code review filters lower your catch rate. Does your review prompt say “only report high-severity issues” or “be conservative”? Opus 5 follows that more literally than earlier models. It finds the bugs, but does not report them, because you told it to be selective. Precision goes up and recall goes down, and it looks like a regression when it is obedience.
Anthropic’s own fix: ask for everything, with a confidence and severity estimate attached, and filter in a separate pass.
New behaviour you need to tame
Three behaviours that are not bugs but do need steering. The wordings below come from Anthropic’s prompting guide.
It writes longer. Not only in conversation, but in the files it writes to disk. And note: the effort dial is not your lever here. Turning it down changes the reasoning, not the length of the visible answer. You do that with an instruction: keep responses focused and brief, keep caveats short, and give a high-level summary unless depth is specifically requested.
It stretches the brief. It sometimes does slightly more than asked, or applies its own judgement about what the task should be without saying so. The instruction that works: deliver what was asked, at the scope intended; make routine judgement calls yourself and check in only when different readings would lead to materially different work; if the request seems mistaken, say so in a sentence and continue with the task as asked.
It delegates too readily. This is a reversal from Opus 4.8, which under-used subagents and needed a nudge to delegate. Opus 5 reaches for them unprompted, and every subagent builds its own context, reports back, and then gets read again by the coordinator. On small jobs that is pure overhead. If your harness supports subagents, put a hard cap on the count and describe when delegation actually pays.
And the wins
So as not to end on warnings only. Three things get easier.
The prompt caching threshold drops from 1,024 to 512 tokens. Short prompts that were just too small to cache now benefit without you changing anything.
You can swap tools mid-conversation without invalidating your cache. Previously, adding or removing a tool meant paying for the entire preceding conversation again. It is in beta, but on long agent sessions that is real money.
And there are automatic fallbacks: if a request is stopped by the safety classifiers, you can have it route to another model automatically instead of getting a refusal back. Useful, and also the kind of mechanism I would want visible in my logs when it fires, because otherwise you quietly receive answers from a different model than you thought.
One last practical note: Opus 5 has its own rate limit pool, separate from the shared Opus 4.x pool. Shifting traffic over neither frees headroom there nor inherits any. Check your limits before you move volume.
That was part four of four. The full series lives here.
Dit stuk verscheen ook in het Nederlands: Vier dingen die breken als je overstapt naar Opus 5.