AI Reasoning Levels Put to the Test: How Claude Builds Websites from a Single Prompt

Tech writer and developer Patrick Campanale has put the reasoning capabilities of Claude 5.5 Opus to a surprising test, challenging the AI model to generate a complete website across every available reasoning effort level using an identical, highly detailed prompt. The experiment, designed to see how varying degrees of internal reasoning impact code generation, asset handling, and overall site design, yielded fascinating results that challenge the assumption that maximum reasoning is always required for high-quality output.

The experiment centered on creating a landing page for Alderwick Estate, a narrative-driven, cozy Victorian countryside adventure modpack for Minecraft 1.21.1 designed by Campanale. To ensure complete fairness and eliminate any potential bias or pre-existing configuration interference, Campanale ran each test in a fresh development session outside of his normal folders. This meant no pre-existing setup files, custom agent instructions, or configuration markers like CLAUDE.md files could influence the model. In fact, to keep the parameters strictly impartial, he had Claude itself write the prompt based on a general description of the project before executing the trials across low, medium, high, extra high, and maximum reasoning effort levels.

Using Claude Code, the model was tasked with generating static web files—including HTML, CSS, and JavaScript with no complex build steps required—and deploying them via Cloudflare Workers once completed. The prompt provided comprehensive creative freedom alongside extensive background lore, technical requirements, art asset references, and installation instructions for prospective players joining the private Minecraft server.

Every Site Got the Same Prompt in a Fresh Session

The core of the experiment relied on consistency. By feeding every reasoning tier the exact same extensive prompt detailing the modpack’s Victorian gothic atmosphere, complex lore featuring three distinct generations of estate owners, regional village dynamics, and custom gameplay mechanics, Campanale sought to isolate how the AI’s internal cognitive allocation changed the final product.

The prompt detailed rich narrative elements, such as the mysterious abandonment of the estate by Lenora twelve years prior, the distinct personalities of neighboring settlements like Alderwick, Bramblemead, and Ironmere, and precise technical instructions for players installing the NeoForge-based modpack via CurseForge. With all these moving parts handed to the model in a blank-slate session, the differences began to emerge primarily in execution time, token consumption, and minor stylistic flourishes.

Low Reasoning Is Fast, and You Can Somewhat Tell

Starting at the bottom of the spectrum, the low reasoning effort tier predictably delivered the fastest and most economical results. Completing the task in just one minute and thirty-two seconds, the low-effort run consumed approximately 10,600 output tokens—roughly a fifth of the token count utilized by any of the higher tiers.

Feedback from colleagues and family members who reviewed the resulting sites indicated that while the low-reasoning output was not anyone’s favorite, it was universally regarded as a fully passable, functional landing page. It stood out as the lightest site of the group, utilizing a fraction of the code of its counterparts. Interestingly, the low reasoning tier also possessed a unique quirk: it was the only effort level that did not independently preview or open the website to verify its appearance before concluding the task. All higher reasoning levels automatically performed this verification step without explicit prompting. Despite the rapid turnaround, Campanale noted that the structural foundation provided by the low-effort run offered an exceptionally solid starting point for anyone looking to quickly spin up a project.

Medium Through Max Land Closer Together Than You’d Expect

Moving up the reasoning scale, the technical resource consumption scaled upward significantly before plateauing into a tight cluster. The medium reasoning effort level required six minutes and thirty-five seconds and used about 43,000 output tokens. Stepping up further, the high reasoning tier took eight minutes and twelve seconds with roughly 50,900 tokens, while the extra high tier finished in seven minutes and twenty-eight seconds utilizing 47,600 tokens. At the ceiling of the capability curve, the maximum reasoning effort level took nine minutes and twenty-five seconds and consumed 56,700 tokens.

Despite the variations in generation time and computational cost, the visual and functional outcomes across the medium, high, extra high, and maximum tiers shared a remarkably high baseline of quality. While maximum reasoning ultimately emerged as the clear favorite among reviewers, the differences were subtle enough that Campanale stated he would feel entirely satisfied using the output from any of the tested tiers without performing any post-generation cleanup or manual code edits.

Every single tier successfully captured the moody, Victorian-inspired aesthetic of the Minecraft modpack, integrating the provided graphical assets and lore elements into cohesive, responsive web pages. Campanale observed that each reasoning level contributed unique charm elements—ranging from clever micro-copy choices in the low tier to distinct font styling in the medium version, decorative design elements in the high and extra high builds, and polished execution across the board in the max version.

The experiment ultimately challenges the notion that developers must always max out an AI model’s reasoning capabilities to achieve production-ready web development results. By demonstrating that even lower effort tiers can produce clean, functional, and aesthetically pleasing sites from complex prompts, the test highlights the remarkable flexibility and reliability of modern AI coding agents in handling creative briefs.