Claude Opus 5.5 Reasoning Levels and Website Generation Efficiency
Testing how varying AI reasoning efforts impact speed, cost, and final website quality.
1Testing AI Reasoning Levels
Patrick Campanale conducted an experiment using Claude Code to build a website for a Minecraft modpack called Alderwick Estate. He tested five distinct reasoning levels of Opus 5.5, ranging from low to max. Each session ran in a fresh environment to ensure no prior configuration files influenced the generated output.
2Low Effort Performance Metrics
The low reasoning level completed the website in just 92 seconds, utilizing approximately 10,600 output tokens. This level was the fastest and most cost-effective option in the test. Notably, it was the only configuration that did not automatically verify the website by opening it before declaring the task finished.
3Scaling Reasoning and Costs
Higher reasoning levels required significantly more time and token resources. Medium reasoning took 6 minutes 35 seconds with 43,000 tokens, while max reasoning reached 9 minutes 25 seconds using 56,700 tokens. These results demonstrate a clear linear correlation between the depth of reasoning effort and the total computational cost incurred.
4Comparing Output Quality
Feedback from colleagues and family indicated that while no one chose the low-effort site as their favorite, it remained a functional and passable product. The max reasoning level was identified as the clear winner in terms of design and polish. However, all versions produced usable results without requiring any additional manual prompting.
5Subjective Design Preferences
Each reasoning level produced unique aesthetic elements that appealed to the author. The low level offered strong wording, while the medium level featured superior header fonts. Higher levels like high and extra-high contributed specific design flourishes, such as unique footer icons and scalloped layouts, showing varied creative outputs across all settings.
6The Role of Prompting
The author utilized a prompt generated by Claude itself to maintain fairness across all test iterations. This strategy suggests that the quality of the initial instruction is a critical factor in AI development. Users can achieve high-quality results by refining their prompts with the model before initiating the actual build process.
7Efficiency Versus Perfection
The experiment concludes that users do not necessarily require max reasoning levels to generate a high-quality, functional website. While max settings provide the most refined results, the lower tiers offer a highly efficient starting point. Developers can effectively balance speed and cost depending on the specific requirements of their web project.
8Implications for AI Development
This test highlights that Claude Opus 5.5 is capable of producing professional-grade code across all reasoning tiers. The findings challenge the assumption that maximum compute is always required for simple web tasks. Future workflows may prioritize lower reasoning levels for rapid prototyping, reserving higher efforts only for complex, multi-layered application architectures.
Original article: How-To Geek — Claude Opus 5.5 Reasoning Levels and Website Generation Efficiency