For two years the standard advice to anyone getting poor results from a language model was to add more instructions. Anthropic spent part of this week telling its customers to start taking them out.
The company published guidance in its documentation on Nov 14 for Opus 5.5, its newest model. According to that documentation, the model produces output more than 30% faster than Opus 5 and spends fewer tokens on the same task. Most prompts written for Opus 5 work unchanged, which is the part customers wanted to hear.
The part they did not expect was the warning attached to it. Prompts that stack constraints inherited from earlier models, or that repeatedly instruct the model to reason step by step, slow Opus 5.5 down and raise the cost of running it. The scaffolding that once helped is now overhead.
Anthropic’s recommendation is to break long instructions into clear acceptance criteria. Say less about the process the model should follow and more about how the result will be judged. A prompt that describes what a good answer looks like, in the terms a reviewer would use, is worth more than one that narrates the reasoning path to get there.
The advice lands on a habit that formed for good reasons. Earlier generations of models did better when they were told to work through a problem in sequence, and an entire vocabulary grew up around that instruction: think step by step, show your work, consider the alternatives before answering. For a while those phrases measurably improved output, and teams wrote them into templates that then propagated across products.
What changed is that the reasoning is now inside the model rather than outside it. A model trained to work through a problem before answering does not need to be asked twice, and asking twice costs tokens on every call. At the scale of a customer-service system or an agent that runs thousands of steps a day, that repetition is a line item.
Anthropic put a number on the speed difference and left the cost side qualitative. More than 30% faster output is the kind of figure that matters to a product team arguing about whether to upgrade: latency is a feature, and a third of it returned is easier to sell internally than a benchmark score.
Analysts said the note is less about prompt craft than about the economics of model upgrades. Every release resets the assumptions a company built its prompts, its evaluation suite and its cost model around, and the vendors have an interest in making the reset cheap. A model that accepts its predecessor’s prompts without modification removes the largest single reason customers delay an upgrade.
That reason is usually not technical. A prompt in production is surrounded by tests, fallback logic and a set of behaviours that someone signed off on months ago. Changing the model means re-running all of it, and the team that owns the system has no incentive to volunteer for that work unless the new version fails visibly on the old prompts.
Anthropic’s guidance is an attempt to lower that barrier from both directions. Telling customers their prompts still work removes the objection; telling them which habits to drop removes the disappointment that follows when an unchanged prompt behaves differently at speed.
The advice also implies a division of labour. If the model handles the reasoning, the prompt should carry the requirements: what counts as correct, what counts as unacceptable, what the output must contain and what it must never contain. That is closer to writing a specification than to writing an incantation, and specifications are easier to test.
There is a commercial logic underneath. Tokens are the unit Anthropic bills for, and a customer who trims a bloated prompt spends less per call but also, in principle, buys fewer tokens. The company’s willingness to recommend the trim suggests it expects volume to grow faster than per-call spend falls, which is the usual bet in this market.
For developers, the harder question is what to do with the prompts that were tuned by trial and error over many months and now sit in a repository with no author. Nobody can say which line was load-bearing. The safest path is to strip the prompt back to its acceptance criteria and let the evaluation suite decide what needs to return.
Anthropic has not said whether the same guidance applies to its smaller models or to the versions still in preview. The documentation covers Opus 5.5, and the company has been consistent in treating each release as its own set of instructions rather than one policy for the family.
The practical effect of a note like this is slow and cumulative. Teams that read it will shorten their prompts, find that most of the deletions cost nothing, and carry a smaller template into the next upgrade. Teams that do not will keep paying for instructions the model stopped needing, and will blame the model when it runs slower than the release notes promised.


