Anthropic has made Claude Opus 5.5 the default model on all paid plans.Alongside this, a guide has been published describing which prompting habits no longer make sense with the new model.The two key points: formulas such as "Think hard" should be removed, and for the thinking effort – the so-called effort level – you should start at the medium level instead of reflexively choosing the highest.
That sounds like a detail question, but it affects every company that uses language models productively: in software development, in editorial processes or in automated workflows.Prompts are rarely disposable there.They sit in system instructions, in agent workflows, in text modules used by teams – and they travel along for years, often unchanged.
Why old magic formulas do not last forever
Phrases such as "Think hard" or "Think step by step" became established because they produced measurably better results with earlier model generations.They were workarounds: the model had to be nudged linguistically into taking a longer solution path.
Newer models, however, control this effort themselves or via dedicated parameters.As a result, the formula loses its purpose – and in the worst case can even interfere, because it gives the model an instruction that competes with the actual task description.The consequence is unpleasant for teams: a prompt that demonstrably helped a year ago is at best neutral today.
Medium instead of maximum
The second tip concerns the setting of the thinking effort.Anthropic recommends starting at the medium level.In practice, many teams choose the highest setting out of caution, following the principle: more thinking cannot hurt.
This assumption is questionable for two reasons.First, more thinking effort costs more computing time and therefore more money as well as longer response times – noticeable in agents that work through a task in many individual steps.Second, high effort is no guarantee of quality: for clearly defined tasks such as rephrasing a text, mapping a field or a simple code change, it brings little.The sensible order is therefore: start with medium effort, check the results, and only turn it up where the quality is actually insufficient.
What this means for existing prompt inventories
Anyone who has integrated language models into production systems should use this model change as an occasion for a stocktake.Experience shows that in larger organisations nobody knows completely where prompts are stored.
- Record the inventory: Where are system prompts, templates and agent instructions – in the code, in the content management system, in automation platforms, in individual employees' notes?
- Sort out old formulas: Instructions that were only meant to control thinking effort can be removed.Substantive specifications regarding role, format, target group and limits remain.
- Set the effort level deliberately: Not across the board, but depending on the task type – and documented so that the decision remains traceable.
- Cross-check: A small, fixed collection of typical tasks with expected results quickly shows whether a change helps or harms.
Model changes belong in the operations plan
That a provider swaps out the default model is convenient and risky at the same time from a user's perspective.Convenient, because a newer model generation is available without any action of your own.Risky, because the behaviour of applications can change without a single line being touched in your own code.
For production systems, it is therefore advisable to specify the model version used explicitly instead of silently staying on the default.That way your own team decides when a change takes place – and can test it before customers see the effects.In parallel, a fixed point in the maintenance cycle is needed at which model versions and provider recommendations are reviewed.Prompts are therefore not text, but configuration: versioned, tested, documented.
Keep prompts lean
One practical lesson from the guide applies regardless of the provider: the less ballast a prompt contains, the easier it is to transfer to new models.Instructions that solely compensate for a weakness of a particular model generation inevitably become outdated.Descriptions of the task, the framework conditions and the desired output format, on the other hand, remain stable.
Anyone who divides their prompts accordingly – the substantive core on one side, model-specific fine-tuning on the other – will only have to touch a small part when future changes occur.
Classification for web and software projects
For projects in which AI generates content, prepares data or supports development tasks, the effort shifts from one-off prompt formulation to ongoing maintenance.Anyone who treats model version, effort level and prompt texts as versioned configuration and safeguards them with a small test collection will survive model changes without a drop in quality.
And anyone who deliberately chooses the amount of reasoning effort per task instead of maximising it across the board keeps response times and operating costs low where the additional depth brings no added value anyway.