10 Things Builders Actually Made With Claude Opus 5.5 in Its First Day

Update — September 25, 2026: New build log — 5 friendslop browser games built with Claude Opus 5.5 (boxing, co-op wizards, karts, paint shooter, sword brawler).

Claude Opus 5.5 launched September 22, 2026, and builders started publishing what they'd made with it within hours. This is a curated list of 10 specific, sourced examples — from Anthropic's own official demo reel to independent projects on X and Hacker News — chosen for having an actual viewable artifact behind the claim, not just a screenshot of a chat.

TL;DR

1. A formally verified production SDK, zero unproven gaps

The most technically substantial item on this list by a wide margin. Boris Cherny used Opus 5.5 with Lean 4 and TLA+ to formally model and verify the Claude Agent SDK, starting from what he describes as "a couple short prompts." Result: 1,529 theorems across 6 Lean models with zero unproven sorry placeholders, 24 bugs found (19 directly by the proofs), and 16 pull requests — six of which removed more code than they added. Cherny doesn't know Lean or TLA+ well himself; the model did the formal modeling.

The reply above is the infographic Cherny had Opus generate summarizing the project's own numbers — pull requests, bugs found, theorem count, and the added/removed line breakdown by PR category.

2. A reverse CAPTCHA

Builder Izzy handed Opus 5.5 an idea — a CAPTCHA that inverts the usual human/bot verification logic — and let it build the implementation directly, posting the working result on X. Izzy's own framing afterward matched Anthropic's "hand over the whole task" prompting advice almost exactly: "I thought the answers would be obvious. They really weren't."

3. Vibe-coded game clones, benchmarked against three other models

Developer Senko ran the same detailed prompts — build a Minecraft-style voxel game, build a Warcraft-style RTS — across Opus 5.5, Fable 5.1, and GPT-6 Astra, and published all six resulting games live for direct comparison at senko.net/vibecode-bench. Cost and time were disclosed alongside the results: Opus 5.5's builds took roughly 45 minutes and cost between $11 and $14 each at xhigh effort via Claude Code on a Max subscription. Commenter kitbrennan noted Astra developed its own distinct visual style across both games, while Claude's versions tracked closer to the source material's original gameplay depth.

4. Recreating Apollo 8's Earthrise photograph from public data

Part of Anthropic's own official launch demo thread: a recreation of the exact moment Apollo 8's famous Earthrise photograph was captured, built from image cues and publicly available mission data rather than the photograph itself as a direct reference. Anthropic didn't publish the underlying prompt or methodology, but it's a genuinely different kind of demo than the usual code-generation showcase — closer to historical reconstruction from partial evidence.

5. A napkin-styled, pen-and-paper coding interface

Designer Kevin Ngo built a Claude Code interface styled entirely as hand-drawn paper and pencil sketches — new sessions open as literal napkins, projects appear as index cards "on the table." Featured directly in Anthropic's official demo thread as one of four highlighted builds.

6. A short story about a watermelon

Also from Kevin Ngo, also featured in Anthropic's official thread: a short story generated with a watermelon as its subject, offered with no further framing by Anthropic beyond the video itself. Included here specifically because it's the one item on this list demonstrating pure creative writing rather than code or visual generation — a useful reminder that Opus 5.5's rewritten communication style applies to creative output, not just technical explanations.

7. A physics-accurate pencil sketch brought to life

The fourth item in Anthropic's official demo thread: a hand-drawn pencil sketch of a catapult-style mechanism, animated so that pulling back the arm and releasing it sends a ball flying with real counterweight physics rather than a simplified or scripted animation. Anthropic didn't specify the underlying implementation (canvas, WebGL, or otherwise).

8. A one-shot website footer, iterated soundtrack

Builder DreW posted a project "made with Claude Opus 5.5," with a specific detail worth noting: a site footer that another builder, Prasenjit Sarkar, separately confirmed was "designed & developed in literally one shot." DreW also iterated the accompanying soundtrack directly with Opus 5.5 across roughly three to four attempts, instructing it to use real instrument sounds rather than synthesized ones — a concrete example of the "add to a running task" mid-session steering pattern Anthropic's own playbook recommends.

9. A four-model, one-prompt design-to-HTML comparison

Builder jjcm ran the identical design brief — a specific image mockup, converted to a working HTML implementation — across Opus 5.5, GPT-6 Astra, Grok 4.7, and MiMo, publishing all four resulting sites live for direct comparison: Opus 5.5's build, Astra's build, Grok 4.7's build, and MiMo's build, all against the same source design. jjcm's own read: Opus 5.5 followed the source design closely and correctly, but ranked just below Astra on overall polish and page-transition handling, placing it "solidly second place" among the four in that specific test — a useful counterweight to purely benchmark-driven comparisons, since this is a single evaluator's direct visual judgment on identical inputs.

10. A code-review harness that beat Fable 5.1 on accuracy and cost

Not flashy, but the most rigorously quantified item on this list. Hacker News's gwd ran a real patch-review harness across multiple models on the same set of code patches: Opus 5.5 found 8 of 14 known issues for $15.40 total; Fable 5.1 found 7 of 14 for $66.34. Included here specifically because it's a genuine production-style workload with disclosed real dollar costs, not a curated demo — the kind of result worth weighing more heavily than any single flashy showcase.

What the spread across this list actually shows about the model

Looking at these 10 examples together rather than individually surfaces a pattern worth naming directly: the technical distance between item 1 (formal mathematical verification of a production codebase) and item 6 (a short story about a watermelon) is enormous, and both were produced by the same model within the same 24-hour window. That range is itself informative about what "general-purpose" actually means in practice for a frontier model in late 2026 — not that it's equally strong at everything, but that the same underlying system is being pointed at genuinely disparate problem types by different users with no apparent friction in switching between them. A formal-methods researcher and a hobbyist building a napkin-styled UI are drawing on the identical model, not specialized variants tuned for their respective domains.

It's also worth noting what's conspicuously absent from a list built purely from what got publicly shared in the first 24 hours: quiet, unglamorous production usage — the bulk of what any newly launched model actually gets used for in its first days — by definition doesn't generate a shareable demo video or a benchmark chart, and so doesn't appear on lists like this one at all. That's a structural bias worth being explicit about: a "10 things people built" list necessarily overrepresents flashy, demo-friendly output relative to the far larger volume of mundane, successful, un-tweeted work happening in parallel — code review, data extraction, routine debugging — none of which produces content suited to a list like this one, even though it likely represents the majority of Opus 5.5's actual usage in that same 24-hour window.

Honest limitations

• This list is curated from public, self-reported posts, not independently reproduced by explainx.ai — where a linked artifact exists (a live site, a GitHub repo, a benchmark page), it was checked to confirm it exists and roughly matches the claim, but underlying costs, timings, and specific prompts are taken from each builder's own account.
• Items 4, 5, 6, and 7 come from Anthropic's own official marketing thread, which is inherently curated to present the model favorably — treat them as demonstrations of capability under ideal, company-controlled conditions, not necessarily representative of typical first-attempt results.
• This is not an exhaustive list — it's a selection chosen specifically for having a checkable, linked artifact behind the claim, from among a much larger volume of Opus 5.5 posts circulating in its first 24 hours.
• Items 2 (Izzy's reverse CAPTCHA), 8 (DreW's footer/soundtrack project), and 10 (gwd's patch-review harness) are not embedded above — no verifiable, stable post URL was available for these three at the time of writing (gwd's account is a Hacker News comment, not a tweet, and has no separate post of its own to link). They're included on the strength of the direct quotes and figures reported in each case, not a linked artifact.

What this means for builders

The spread across this list is itself informative: formal verification, game development, creative writing, physics simulation, design-to-code translation, and code review all appear here from the same model within its first day — which tracks with Anthropic's own framing of Opus 5.5 as a broadly capable daily driver rather than a narrowly specialized tool. If you're deciding what to try first, items 1 and 10 — the formal verification project and the code-review harness — are the most useful models for real production workflows specifically because they come with disclosed, checkable numbers rather than just a polished demo video.

Related on explainx.ai

• Update — September 25, 2026: A separate, independent demo went even more viral — a single prompt reportedly produced a 2:16 animated "Western civilization" video by having Claude write its own animation code: Claude Opus 5.5's one-shot Western civilization video, explained →
• Claude Opus 5.5 Launch: Every Benchmark and Reaction
• Boris Cherny Used Opus 5.5 to Formally Verify the Claude Agent SDK
• How to Actually Use Claude Opus 5.5: Anthropic's Own Prompting Playbook
• Claude Fable 5.1 vs Claude Opus 5.5: Which One Do You Actually Need

Primary sources: @claudeai on X, September 22-23, 2026, and individual builders' own posts, linked above.

This post catalogs publicly shared demos as of September 23, 2026. Costs, timings, and outcomes are as reported by each builder.