What Anthropic published
Anthropic has released a prompting guide for Claude Opus 5.5, the model it launched on September 22. 2 The guide makes two main recommendations. Developers should test several effort levels instead of carrying over the setting they used with Claude Opus 5. Chat applications should also consider removing system-prompt lines that tell Claude to think carefully before it answers. 2
The guide is not a call to rewrite everything. Anthropic says prompts written for Opus 5 should still work well without changes, and that its existing Opus 5 guidance is a reasonable place to start. 2 The advice is aimed at tuning, not at fixing breakage.
Effort settings are the headline concern
Two outlets covered the guide, and they frame it differently. AI Weekly's coverage puts cost first. Its headline says Anthropic warns that old Opus 5 effort settings now "blow out tokens" on the new model. It files the story under coding tools and prompt engineering. 1 Search Engine Journal's report is more measured. It describes the advice as a request to retest effort levels and does not stress runaway token use in the material it presents. 2
The two accounts can both be true. Search Engine Journal confirms the core instruction: test effort levels again instead of reusing old ones. 2 AI Weekly's emphasis on token consumption suggests the reason behind that instruction: a setting that was well calibrated for Opus 5 may produce more output, or more internal reasoning, on Opus 5.5. 1 The coverage here does not give figures for how much token use changes, so the size of the effect is unclear. Developers should measure it on their own workloads rather than assume a particular multiplier.
This matters for cost. Teams running Claude in production, especially in coding tools where long agentic sessions are common, often set an effort level once and then stop thinking about it. If the same setting now uses noticeably more tokens, bills and latency could rise quietly after a model swap, even when output quality looks unchanged. Anthropic's advice to run a fresh comparison across effort levels works as an early warning about that kind of silent cost drift.
Dropping "think carefully"
The second recommendation targets a habit common in chat prompts. Many system prompts include a line asking the model to think carefully, step by step, or deliberately before replying. Anthropic tested this in a chat product. Removing the think-carefully line made replies start sooner, and the company reported "no clear decline in the quality of the reply." 2
The finding fits a wider pattern in how frontier models are now used. Instructions like "think carefully" were added when models gained clear benefits from being pushed to reason more. Newer models expose reasoning depth through explicit controls such as effort levels, so a freeform instruction may now duplicate that control or push the model into extra deliberation it doesn't need. In a chat interface, where time to first response strongly shapes how users perceive quality, that extra deliberation has a real cost.
The wording of Anthropic's claim is limited. "No clear decline" is not the same as "no decline," and it comes from one internal test in one chat setting. 2 Teams with specialized tasks, such as complex analysis, multi-step math, or high-stakes coding, may still find that an explicit reasoning nudge helps. The safer reading is that the line is now worth A/B testing instead of keeping by default.
How to read it
Taken together, the guide treats Opus 5.5 as a model whose defaults and sensitivities have shifted enough that inherited configurations deserve a second look. It does not require migration work. The prompts themselves are expected to carry over. 2 The settings and habits around them are what need checking.
The difference in framing between the two outlets is mostly about emphasis. AI Weekly leads with the financial risk of doing nothing. 1 Search Engine Journal leads with the practical steps Anthropic recommends. 2 For most developers, cost is the more urgent concern. If a team upgrades to Opus 5.5 and keeps its old effort level, that is the scenario most likely to produce an unwelcome surprise, and AI Weekly's coverage suggests the surprise would show up in token counts. 1
The practical takeaway is short. After switching to Opus 5.5, benchmark a few effort levels against real tasks, track both token use and output quality, and test whether "think carefully" lines in chat system prompts are still helping. Anthropic's own data point suggests they may only be slowing replies down. 2
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.