Claude Opus 5.5 in the Wild: One-Hour Award Site, 3D Space Voyage and a Single Dashboard

The video opens with a bold claim for Opus 5.5 as the strongest model you can get online right now, beating Fable 5.1 and GPT Astra 6 on several benchmarks while running about 30% faster and costing 40% less than the prior Opus generation on typical workloads. Anthropic’s official page frames it similarly: Opus 5.5 keeps Fable 5.1-level work quality while cutting the typical bill versus Opus 5 by 40%, with list prices moving from $5 and $25 per million tokens to $4 and $20, and cache reads falling from $0.50 to $0.20.

The host quickly tours 3D showpieces circulating on X. Stefan’s single-prompt Blender scene built and animated from scratch, Higsfield’s Unreal Engine underwater game level, an octopus modeled from zero in Blender with bones and textures then rigged and animated, and Alex’s claymorphism animation made in Blender in one go are shown in sequence. The point is a clear jump in 3D asset generation, even if that is not where the host wants to stay.

The focus shifts deliberately to web design. The host wants to reuse the same prompts he runs on every model to see if Opus 5.5 can deliver real user experience: micro-animations, GSAP or 3JS interactions, and professional-grade typography and color hierarchy. The question is not whether it can render a demo, but whether it can craft an actual experience.

The first big test runs inside the Clawed app: an interactive, award-style site about Opus 5.5 itself, with GSAP/3JS and micro-animations required and a narrative that teaches the model as you scroll. The run lasts a full hour, much longer than prior similar generations that finished in minutes. The resulting page is framed as ‘Opus that never stops thinking,’ showing $4 input and $20 output per million tokens, 20% below the prior Opus, thinking that cannot be toggled off, current medium equaling old high, cache reads down 60%, agent coding using about half the tokens, and less verbose outputs. Cards and charts carry those claims, buttons bounce with micro-interactions, and the typography and color hierarchy are judged as genuinely good, with the previously noisy documentation trimmed.

The second test is a redesign of an existing site, typeUI.sh, in the ‘design UI with AI’ lane. The brief asks for better micro-animation, clearer visual hierarchy and a touch more Framer Motion-style liveliness to create a more stimulating page. This run also lasts about an hour, building piece by piece. Keeping the original’s analytics and intro blocks, the new version adds a ‘Build better UI with AI’ hero, one-click color-way switching labeled ‘Same prompt, different taste,’ selectable design skills, clickable UI prompt cards, animation and toggle switches, model-generated noise textures, brand kits and UI audit sections, plus analytics. Every section has hover and transition polish, which the host compares to work expected from a professional graphic designer.

The third test moves to full 3D for a space agency website, a classic prompt the host uses across many models and one where Opus 5.5 is expected to shine. It finishes in 51 minutes. Opened in the browser, it starts with ‘Build on Earth, bound for anywhere else’ over an Earth-and-Sun visual, then as you scroll the Earth expands with topographic detail and scroll-jacking, a rocket is assembled in 3JS on screen, launches from the pad and blasts into orbit, and the scroller carries you through planets. The host calls it up to 10 times more immersive than his prior Astra 6 run on the same prompt, feeling like an actual journey outward.

The fourth test targets tool use and SDK/MCP calls. The goal is to consolidate scattered daily checks — email, YouTube analytics, Webflow/Wix contact forms — into a single dashboard instead of visiting each app manually. The build uses Zapier’s MCP layer, described as a governed bridge to 9,000+ apps and 40,000+ actions via the Model Context Protocol, with SOC 2 certification, OAuth handling and dynamic tool discovery working with Claude, Codex, Cursor and similar clients, plus TypeScript and Python SDKs. After connecting Notion, Gmail, YouTube, Webflow, Wix and Discord and polishing with a bit of extra chat, the resulting board shows 26 emails, 100k views, a down-trending subscriber count, calendar entries, recent video activity and connected website status with light and dark themes. The host gets excited about extending it to Slack threads and to Hermes and OpenClaw agent interactions in one central place.

The time-price-quality triangle becomes the through-line. Five prompts cost about five hours because the model thinks before it builds. It is not slow in the naive sense, but deliberately long-horizon, like handing work to a person and waiting. The payoff each time is described as stunning, and the host feels Astra-class models have not matched this single-prompt quality yet, making patience worth it. That split separates quick-draft users from deep-craft teams.

Official context helps ground the discounts and technical notes. Anthropic says Opus 5.5 needs less compute to serve, so at default settings it costs 40% less than Opus 5 on typical work, cache reads dominate agentic and coding costs so the $0.20 read price matters most, output generation is more than 30% faster, context is 1M with 128K max output, and the automated behavioral audit scores are the strongest to date. Because capability approaches Fable 5.1 in biology and cybersecurity, deployment uses similar safeguards and vetted access programs. Four breaking changes are documented: thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to model and conversation, and text between tool calls can appear in thinking blocks and go quiet when streamed unless a display value is set — matching the host’s observation that longer polishing needs extra chat.

What does this mean for creators and builders? For web teams that need real micro-interaction and hierarchy rather than a minute-scale draft, an hour of patience is a rational trade. Early testers report a 680,000-line code migration done in a day and a web-app load-time cut succeeding on 39 of 40 pages while preserving behavior, compared to smaller or behavior-changing gains on Opus 5. A no-code dashboard that unifies all accounts via Zapier MCP promises daily relief for creators juggling many inboxes. Both benefits, however, hinge on adapting to the one-hour-per-prompt rhythm and on verifying that the 40% saving materializes on cache-heavy workloads in your own profile; otherwise the advantage stays on paper.